U.S. patent application number 13/961534 was filed with the patent office on 2013-12-12 for modular digital signal processing circuitry with optionally usable, dedicated connections between modules of the circuitry.
This patent application is currently assigned to Altera Corporation. The applicant listed for this patent is Altera Corporation. Invention is credited to Suleyman Sirri Demirsoy, Martin Langhammer, Wai-Bor Leung, David Lewis, Yi-Wen Lin, Henry Y. Lui, Volker Mauer, Keone Streicher, Hyun Yi.
Application Number | 20130332497 13/961534 |
Document ID | / |
Family ID | 42679185 |
Filed Date | 2013-12-12 |
United States Patent
Application |
20130332497 |
Kind Code |
A1 |
Streicher; Keone ; et
al. |
December 12, 2013 |
MODULAR DIGITAL SIGNAL PROCESSING CIRCUITRY WITH OPTIONALLY USABLE,
DEDICATED CONNECTIONS BETWEEN MODULES OF THE CIRCUITRY
Abstract
Digital signal processing ("DSP") circuit blocks are provided
that can more easily work together to perform larger (e.g., more
complex and/or more arithmetically precise) DSP operations if
desired. These DSP blocks may also include redundancy circuitry
that facilitates stitching together multiple such blocks despite an
inability to use some block (e.g., because of a circuit defect).
Systolic registers may be included at various points in the DSP
blocks to facilitate use of the blocks to implement systolic form,
finite-impulse-response ("FIR"), digital filters.
Inventors: |
Streicher; Keone; (San
Ramon, CA) ; Langhammer; Martin; (Alderbury, GB)
; Lin; Yi-Wen; (Glendale, CA) ; Leung;
Wai-Bor; (Milpitas, CA) ; Lewis; David;
(Toronto, CA) ; Mauer; Volker; (Lacey Green,
GB) ; Lui; Henry Y.; (Millbrae, CA) ;
Demirsoy; Suleyman Sirri; (London, GB) ; Yi;
Hyun; (San Jose, CA) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Altera Corporation |
San Jose |
CA |
US |
|
|
Assignee: |
Altera Corporation
San Jose
CA
|
Family ID: |
42679185 |
Appl. No.: |
13/961534 |
Filed: |
August 7, 2013 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
12716878 |
Mar 3, 2010 |
8549055 |
|
|
13961534 |
|
|
|
|
12380841 |
Mar 3, 2009 |
|
|
|
12716878 |
|
|
|
|
Current U.S.
Class: |
708/203 |
Current CPC
Class: |
H03H 17/06 20130101;
G06F 7/5324 20130101; H03H 2220/06 20130101; G06F 17/10 20130101;
G06F 2207/382 20130101; H03K 19/17732 20130101 |
Class at
Publication: |
708/203 |
International
Class: |
G06F 17/10 20060101
G06F017/10 |
Claims
1. A digital signal processing ("DSP") circuit block comprising: a
first multiplier circuit that can perform a first N bit by N bit
("N.times.N") multiplication; a second multiplier circuit that can
perform a second N.times.N multiplication; first shifter circuitry
for shifting outputs of the first multiplier circuit by a
selectable one of (1) zero bit positions and (2) N bit positions
toward greater arithmetic significance; first compressor circuitry
for additively combining outputs of the first shifter circuitry and
the second multiplier circuit; circuitry for routing outputs of the
first compressor circuitry to a first other DSP circuit block that
is adjacent to the DSP circuit block; and second compressor
circuitry for additively combining outputs of the first compressor
circuitry and outputs received from a second other DSP circuit
block that is adjacent to the DSP circuit block.
2. The DSP circuit block defined in claim 1 further comprising:
adder circuitry for additively combining outputs of the second
compressor circuitry.
3. The DSP circuit block defined in claim 2 wherein the adder
circuitry comprises carry-propagate adder circuitry.
4. The DSP circuit block defined in claim 2 further comprising:
first register circuitry for registering outputs of the adder
circuitry.
5. The DSP circuit block defined in claim 1 further comprising:
first, second, third, and fourth pluralities of data signal input
leads for supplying inputs for processing by the multiplier
circuits; and pre-adder circuitry for additively combining the
inputs from at least two of the pluralities of data signal input
leads.
6. The DSP circuit block defined in claim 5 further comprising:
circuitry for selecting outputs of the pre-adder circuitry as an
alternative to inputs from at least one of the pluralities of data
signal input leads for processing by at least one of the multiplier
circuits.
7. The DSP circuit block defined in claim 1 further comprising:
selectively usable systolic register circuitry in input data signal
paths to one of the multiplier circuits.
8. The DSP circuit block defined in claim 1 further comprising:
selectively usable systolic register circuitry in output data
signal paths for data output by the second compressor
circuitry.
9. The DSP circuit block defined in claim 1 further comprising:
circuitry for registering signals indicative of outputs of the
second compressor circuitry; and circuitry for selectively applying
outputs of the circuitry for registering to inputs of the second
compressor circuitry for additive combination by the second
compressor circuitry with the outputs of the first compressor
circuitry.
10. The DSP circuit block defined in claim 1 wherein the circuitry
for routing comprises: second shifter circuitry for shifting the
outputs of the first compressor circuitry by a selectable one of
(1) zero bit positions and (2) N bit positions toward lesser
arithmetic significance.
11. The DSP circuit block defined in claim 1 wherein the circuitry
for routing comprises: alternative routing circuitry for routing to
the first other DSP circuit block signals output by circuitry for
selectively routing in the second other DSP circuit block.
12. The DSP circuit block defined in claim 1 wherein the circuitry
for routing comprises: alternative routing circuitry for routing to
the first other DSP circuit block signals indicative of outputs of
the second compressor circuitry.
13. The DSP circuit block defined in claim 12 wherein the
alternative routing circuitry comprises: systolic register
circuitry that is selectively usable for the signals indicative of
outputs of the second compressor circuitry.
14. The DSP circuit block defined in claim 1 wherein the second
multiplier circuit can alternatively perform (1) a 0.5N-bit by
M-bit ("0.5N.times.M") multiplication, and (2) a 0.5N bit by N bit
("0.5N.times.N") multiplication, where M is equal to 1.5N.
15. The DSP circuit block defined in claim 14 further comprising:
circuitry for selectively routing bits of an M bit multiplicand
data word and an M bit multiplier data word to the first and second
multiplier circuits so that the first multiplier circuit can
multiply the N most significant bits of the multiplicand data word
by the N most significant bits of the multiplier data word to
produce a first partial product data word, and so that the second
multiplier circuit can multiply (1) the M bits of the multiplicand
data word by the 0.5N least significant bits of the multiplier data
word to produce a second partial product data word, and (2) the N
most significant bits of the multiplier data word by the 0.5N least
significant bits of the multiplicand data word to produce a third
partial product data word.
16. The DSP circuit block defined in claim 15 wherein the second
multiplier circuit comprises: third compressor circuitry for
additively combining the second and third partial product data
words, with bits of the third partial product data word being
increased in arithmetic significance by 0.5N bit positions relative
to bits of the second partial product data word.
17. The DSP circuit block defined in claim 1 wherein the circuitry
for routing is dedicated to only conveying signals between
instances of the DSP circuit block.
18. The DSP circuit block defined in claim 1 wherein the outputs
received from a second other DSP block are received via circuitry
that is dedicated to only conveying signals between instances of
the DSP circuit block.
19-33. (canceled)
Description
[0001] This is a continuation-in-part of application Ser. No.
12/380,841, filed Mar. 3, 2009, which is hereby incorporated by
reference herein in its entirety.
BACKGROUND OF THE INVENTION
[0002] This invention relates to digital signal processing ("DSP")
circuitry, and more particularly to arrays of DSP circuit modules
or blocks that can optionally work together to perform DSP
operations of greater complexity and/or greater mathematical or
arithmetic precision, and that when thus needed to work together,
can accommodate the possibility that a circuit defect may make a
DSP block unusable.
[0003] Various circumstances may call for an integrated circuit to
be fabricated with multiple instances of blocks or modules of DSP
circuitry. An example of such an integrated circuit is a
programmable logic device ("PLD") or a field-programmable gate
array ("FPGA"). Such a device may have a plurality of rows of
various kinds of circuitry, such as relatively general-purpose
programmable logic. Each such row may also include a block of DSP
circuitry (i.e., circuitry that is hard-wired to at least some
degree to perform a particular DSP function or a particular set of
DSP functions). It can be desirable in such a situation to size the
DSP blocks so that they fit within the (row) boundaries of the
other circuitry in the row. This may mean that a DSP block is too
small, by itself, to perform some DSP functions that it may be
desired for the integrated circuit to perform. In such cases it
would be desirable to facilitate optional "stitching together" of
multiple DSP blocks in various ways so that larger DSP functions
can be efficiently performed in two or more DSP blocks. However, a
countervailing concern may be that if any portion of the circuitry
associated with DSP blocks that need to be stitched together is not
usable (e.g., because of a manufacturing defect in the integrated
circuit), that can make it much more difficult or impossible to
stitch together those DSP blocks. This may greatly increase the
chances that a partly defective integrated circuit cannot be used
at all.
SUMMARY OF THE INVENTION
[0004] In accordance with certain possible aspects of the present
invention, DSP circuit blocks are provided that can more easily
work together to perform larger (e.g., more complex and/or more
arithmetically precise) DSP operations if desired. For example,
such DSP blocks may include routing circuitry for optionally or
selectively routing signals to other DSP blocks on either side of
each block.
[0005] In accordance with other possible aspects of the invention,
the inter-DSP-block routing circuitry may also include redundancy
capability that enables an outbound signal to go to either of two
other DSP blocks to one side of each DSP block, and that enables an
inbound signal to come from either of two DSP blocks to one side of
each DSP block. If some DSP block that it is desired to stitch to
another DSP block cannot be used (e.g., because of a circuit
defect), this redundancy capability allows the stitching together
of DSP blocks to take place by effectively skipping over the
defective DSP block.
[0006] In accordance with still other possible features of the
disclosure, a DSP circuit block may include two multiplier circuits
("multipliers"). These multipliers may operate independently or at
least partly independently of one another in various ways.
Alternatively, these multipliers may operate together in various
ways. One of these multipliers may have some optionally usable
extended capabilities that can facilitate using the two multipliers
together to perform one multiplication that is significantly larger
than the base multiplication performable by either multiplier
alone.
[0007] In accordance with yet other possible features of the
disclosure, a DSP block may include systolic registers at various
points in the circuitry to help the DSP block perform functions of
a systolic form, finite-impulse-response ("FIR"), digital
filter.
[0008] Further features of the invention, its nature and various
advantages, will be more apparent from the accompanying drawings
and the following detailed description.
BRIEF DESCRIPTION OF THE DRAWINGS
[0009] FIG. 1 is a simplified schematic block diagram of an
illustrative embodiment of representative circuitry in accordance
with the invention.
[0010] FIG. 2 is a simplified schematic block diagram of an
illustrative embodiment of circuitry that can be used in FIG. 1
type circuitry in accordance with the invention.
[0011] FIG. 3 is a more detailed, but still simplified, schematic
block diagram of an illustrative embodiment of a representative
portion of circuitry of the type shown in FIG. 1 in accordance with
the invention.
[0012] FIG. 4 shows the FIG. 3 circuitry rotated clockwise
90.degree. to facilitate some aspects of discussion of that
circuitry.
[0013] FIG. 5 is a simplified diagram illustrating certain aspects
of performance of a particular DSP operation in accordance with the
invention.
[0014] FIG. 6 is a simplified schematic block diagram showing an
illustrative embodiment of DSP circuitry for performing the FIG. 5
operation in accordance with the invention.
[0015] FIG. 7 is a simplified diagram illustrating certain aspects
of performance of another DSP operation in accordance with the
invention.
[0016] FIG. 8 is a simplified schematic block diagram showing an
illustrative embodiment of DSP circuitry for performing the FIG. 7
operation in accordance with the invention.
[0017] FIG. 9 is a simplified diagram illustrating certain aspects
of performance of yet another DSP operation in accordance with the
invention.
[0018] FIG. 10 is a simplified schematic block diagram showing an
illustrative embodiment of DSP circuitry for performing the FIG. 9
operation in accordance with the invention.
[0019] FIG. 11 is a simplified schematic block diagram of another
illustrative embodiment of DSP block circuitry in accordance with
certain further possible aspects of the disclosure.
[0020] FIG. 12 is a simplified schematic block diagram of an
illustrative embodiment of a representative portion of circuitry
like that shown in FIG. 11 in accordance with certain possible
aspects of the disclosure.
[0021] FIG. 13 is a simplified schematic block diagram of an
illustrative embodiment of another representative portion of
circuitry like that shown in FIG. 11 in accordance with certain
possible aspects of the disclosure.
[0022] FIG. 14 is a simplified schematic block diagram of an
illustrative embodiment of still another portion of circuitry like
that shown in FIG. 11 in accordance with certain possible aspects
of the disclosure.
[0023] FIG. 15 is a more detailed, but still simplified, schematic
block diagram of an illustrative embodiment of circuitry like that
shown in FIG. 14 in accordance with certain possible aspects of the
disclosure.
[0024] FIG. 16 is a simplified schematic block diagram of an
illustrative embodiment of still another representative portion of
circuitry like that shown in FIG. 11 in accordance with certain
possible aspects of the disclosure.
[0025] FIG. 17 is a simplified schematic block diagram of an
illustrative embodiment of a further portion of circuitry like that
shown in FIG. 11 in accordance with certain possible aspects of the
disclosure. FIG. 17 shows configuration of the circuitry to perform
a DSP operation like that shown in FIG. 18.
[0026] FIG. 18 is a simplified diagram illustrating certain aspects
of performance of a DSP operation in accordance with certain
possible aspects of the disclosure.
[0027] FIG. 19 shows another illustrative configuration of portions
of the FIG. 17 circuitry in accordance with certain other possible
aspects of the disclosure. FIG. 19 shows configuration of the
circuitry to perform a DSP operation like that shown in FIG.
20.
[0028] FIG. 20 is a simplified diagram showing certain aspects of
performance of another DSP operation in accordance with certain
possible aspects of the disclosure.
[0029] FIG. 21 is a simplified schematic block diagram of an
illustrative embodiment of a still further portion of circuitry
like that shown in FIG. 11 in accordance with certain possible
aspects of the disclosure.
[0030] FIG. 22 is a simplified diagram illustrating certain aspects
of yet another DSP operation in accordance with certain possible
aspects of the disclosure.
[0031] FIG. 23 is a simplified block diagram of an illustrative
embodiment of two instances of circuitry like that shown in FIG. 11
on an integrated circuit in accordance with certain possible
aspects of the disclosure.
[0032] FIG. 24 is similar to FIG. 23, but for three instances of
FIG. 11 type circuitry on an integrated circuit in accordance with
the disclosure.
[0033] FIG. 25 shows a known equation for a known type of digital
filter function.
[0034] FIG. 26 shows known circuitry for implementing the FIG. 25
filter function.
[0035] FIG. 27 shows alternative "systolic" circuitry for
implementing the FIG. 25 filter function.
[0036] FIG. 28 is basically a redrawing of FIG. 27 to better
correlate with other FIGS. herein.
[0037] FIG. 29 shows another way that the functions shown in FIG.
28 can be implemented in accordance with certain possible aspects
of the disclosure.
[0038] FIG. 30 shows two adjacent instances (on an integrated
circuit) of circuitry like that shown in FIG. 21 in accordance with
certain possible aspects of the disclosure.
[0039] FIG. 31 shows an example of a DSP operation that can be
performed in the FIG. 30 circuitry in accordance with certain
possible aspects of the disclosure.
[0040] FIG. 32 shows multiple adjacent instances (on an integrated
circuit) of circuitry like that shown in FIG. 11 in accordance with
certain possible aspects of the disclosure.
[0041] FIG. 33 is a simplified schematic block diagram of
representative portions of illustrative circuitry in accordance
with certain possible aspects of the disclosure.
DETAILED DESCRIPTION
[0042] A representative portion of an illustrative embodiment of an
integrated circuit ("IC" or "device") 10 in accordance with the
invention is shown in FIG. 1. The circuitry shown in FIG. 1
includes representative portions of several representative rows R4
through R9 of circuitry. Any one of these rows may be referred to
generally as row RN.
[0043] Each row RN includes a block of digital signal processing or
DSP circuitry 100. Each row RN also includes areas of other
circuitry 20 to the left and right of the DSP block in that row.
That other circuitry 20 may include such components as logic
circuitry and interconnection conductors for conveying signals to
and from that row's DSP block, and also to, from, and between that
row's logic and other circuitry, as well as between the rows.
Various aspects of some or all of this circuitry may be
programmable, e.g., to enable a generic device 10 to be put to any
of several different uses. For example, although the circuitry of
DSP blocks 100 is typically hard-wired to some extent to perform
certain DSP functions, that circuitry may also be programmable to
some extent so that certain aspects of the DSP operations performed
can be varied by different users of the device to meet each user's
particular needs and requirements. Such programming may be the
result of storing control data in memory cells on the integrated
circuit, blowing fuses on the integrated circuit, mask programming
the integrated circuit, or any other suitable programming technique
or technology.
[0044] Device 10 may be constructed so that the rows are redundant.
This means, for example, that each row RN is identical or
substantially identical to all other rows. In addition, device 10
may be constructed so that if any row RN is defective, the row
immediately above or below that row can effectively take the place
of the defective row. In addition, all other rows above or below
the row that is effectively taking the place of the defective row
effectively replace other adjacent rows. In this way, many devices
10 that are inadvertently manufactured with some defective
circuitry can still be used, which significantly increases the
yield of the manufacturing process. FIG. 1 shows representative
circuitry for achieving this redundancy in the DSP block areas 100.
Additional redundancy circuitry is provided for other portions of
each row RN but is not shown in FIG. 1 to avoid over-complicating
the drawings (and because those other portions of the rows are not
the primary focus of this invention). The DSP block redundancy
circuitry will be described in general terms in the next several
paragraphs.
[0045] In accordance with the present invention, a DSP block 100
can extend its functionality by sending certain signals to and/or
receiving certain signals from an adjacent DSP block or blocks.
These are relatively direct connections between adjacent DSP blocks
100. These inter-DSP-block connections do not rely on other, more
general, interconnection resources of device 10 such as the
interconnection resources that form part of circuitry 20. Rather,
these inter-DSP-block connections go directly from one DSP block
100 to another adjacent DSP block 100, subject only to the
possibility that the redundancy circuitry that will now be
described may be employed to allow these inter-DSP-block signals to
effectively "jump over" a row that has been taken out of service
due to one or more defects in that row.
[0046] Considering representative row R7 as an example, a signal
that needs to go (in accordance with this invention) from the DSP
block 100 in row R7 to another DSP block may originate at node A in
the row R7 DSP block. This signal is applied to one selectable
input terminal of the multiplexer circuitry ("mux") 110 in that DSP
block, and also to one selectable input terminal of the similar mux
110 in the DSP block 100 in the row R6 above row R7. The output
signal of mux 110 in row R7 goes to a destination in the DSP block
100 in row R6. The output of mux 110 in row R6 goes to a
destination in the DSP block 100 in row R5. If row R6 is not
defective and is therefore in use, the mux 110 in row R7 is
controlled to select the signal from node A for application to row
R6. But if row R6 is defective and is therefore effectively cut out
of the circuitry of device 10, then mux 110 in row R6 is controlled
to select the signal from node A for application to row R5. This
example shows how redundancy muxes 110 can be used to apply a
signal from the DSP block 100 in any row to the DSP block 100 in
the row immediately above or to the DSP block 100 two rows above
the source row.
[0047] Redundancy muxes 120 can be used similarly to route a signal
from any DSP block 100 to either the DSP block 100 immediately
below or the DSP block 100 two rows below. For example, a signal
that originates at node B in the DSP block 100 in row R5 is applied
to one selectable input of the muxes 120 in each of rows R6 and R7.
If row R6 is in use, the mux 120 in row R6 is controlled to apply
the signal from node B to the destination in row R6. On the other
hand, if row R6 is not in use, then the mux 120 in row R7 is
controlled to apply the signal from node B to the destination in
row R7.
[0048] FIG. 2 shows illustrative embodiments of how a
representative redundancy mux 110 or 120 may be controlled. Mux
110/120 can select either of its primary or selectable inputs H or
I to be its output signal J based on the logical state of its
selection control input signal K. For example, if K is zero, J may
be H; and if K is one, J may be I. K may come from a fuse (F) 130
on device 10, memory cell (M) 130 on the device, or any other
programmable feature 130 on the device. Such a fuse, memory cell,
or the like 130 may be programmed to the appropriate state for each
such element 130 after it has been determined which (if any) row RN
is defective and must therefore be effectively cut out of device 10
for purposes of normal use of the device.
[0049] A point to be made in connection with FIG. 1 is that the
width W of each row (between adjacent rows) is approximately the
same for (1) the DSP block 100 in that row and (2) the other
circuitry 20 in that row. This facilities efficient use of the area
on device 10 in a context, like the present one, in which each row
is substantially identical to the other rows and in which any row
may be completely taken out of service (if defective) and
seamlessly replaced by another row. When a row is thus taken out of
service and replaced by another row, all functions of the original
row are preferably automatically reassigned to the replacement row.
The user of the device does not need to be concerned with, or even
know, which rows are being used and which row is not being
used.
[0050] A possible consequence of making the height of each DSP
block 100 the same as the height of other circuitry 20 in a row is
that it tends to force the size or functional capacity of each DSP
block to be commensurate with the size and capacity of the adjacent
other circuitry 20 in the row that includes that DSP block. For
example, a DSP block 100 may only be able to get a certain number
of input signals from the adjacent circuitry 20 in the row that
includes that DSP block. This may limit the number and/or size of
the DSP operations (e.g., multiplications) the DSP block can
perform. However, some users of device 10 may want to perform
larger multiplications than can be performed in one such limited
DSP block. The present invention therefore provides for extending
the multiplication and other capacities of one DSP block 100 by,
for example, allowing some portions or aspects of a large
multiplication and/or other DSP operation to be performed in
another adjacent DSP block 100. In accordance with the present
invention, this is done by sending signals substantially directly
between adjacent DSP blocks via redundancy circuitry like that
shown in general at 110 and 120 in FIG. 1. Substantially direct
inter-DSP-block connections are used in this way to avoid the delay
and possible other resource exhaustion that might result from
instead attempting to use the more general-purpose interconnection
resources of other circuitry 20. Redundancy like 110 and 120 is
used in these inter-DSP-block interconnections so that if a row
must be taken out of service, the signals that need to go from one
DSP block 100 to another can bypass the non-functioning DSP block
100 in the row that has been taken out of service.
[0051] FIG. 3 shows an illustrative embodiment of a DSP block 100
in accordance with the invention. The various elements of DSP block
100 will be identified early in what follows. More details
regarding how these elements can be used will be provided later.
FIG. 3 shows representative DSP block 100 in the same orientation
as is shown in FIG. 1. However, because this orientation can be
difficult to talk about when it comes to explaining arithmetic
operations performed by various components of the DSP block, the
substance of FIG. 3 is exactly reproduced in a different
orientation in FIG. 4. In particular, FIG. 4 is rotated clockwise
90.degree. relative to FIG. 3. Thus "left" or the like in FIG. 4 is
the same as "bottom," "down," "below," or the like in FIG. 3;
"right" or the like in FIG. 4 is the same as "top," "up," "above,"
or the like in FIG. 3, etc. FIGS. 1 and 3 make it clear that the
above-mentioned substantially direct connections between adjacent
DSP blocks 100 go from row to row in device 10. Although FIGS. like
FIG. 4 depict a representative DSP block 100 rotated 90.degree.,
the following discussion of FIGS. like FIG. 4 will still refer to
such inter-DSP-block signals going from one "row" to another "row",
even though FIGS. of this kind may make it appear that signals
traveling left or right are going into another column of circuitry
rather than into another row of circuitry.
[0052] Turning now to the basic structure of representative DSP
block 100 as shown in either FIG. 3 or FIG. 4, block 100 includes
two 18 by 18 ("18.times.18") multiplier circuits 200a and 200b.
Each of multipliers 200 can multiply together two 18-bit data words
(represented by electrical data signals from the other circuitry 20
in the row that includes the DSP block 100 under consideration) and
produce two product vectors (again represented by electrical
signals) in redundant format (e.g., one 36-bit sum vector and one
36-bit carry vector (because the multiplier does not include a
final carry-propagate adder ("CPA") for producing a normal final
product value)). Controllable shifter circuit 202 is controllable
to shift the sum and carry signal vectors output by multiplier 200a
either (1) 18 bit positions to the left (increased numerical or
arithmetic significance), or (2) not at all (i.e., no shift to the
left and therefore no change in numerical or arithmetic
significance). Another possible output condition for element 202
may be outputting data signals that are all zeros. As in the case
of FIG. 2, one or more fuses or memory cells like 130 may be
programmed to control element 202 with respect to whether or not to
shift the outputs of multiplier 200a to the left as described above
(or, as a third possibility, to output 0 data as mentioned above).
As another possible alternative, element 202 may be mask
programmable with respect to the function that it performs. In all
cases throughout this specification, "shifting" typically means
routing signals to a different set of electrical leads going to the
next downstream circuit element than the set of electrical leads
that would otherwise be used to convey those signals (if not
shifted) to the next downstream circuit element. The "arithmetic
significance," "bit position," "order," or the like of a bit signal
is typically the result of which of several electrical leads that
signal is on.
[0053] The next element in representative DSP block 100 is
four-to-two ("4-2") compressor circuit 210. Compressor 210 can
combine the two sum and carry signal vectors it receives from each
of shifter 202 and multiplier 200b (i.e., a total of four such
signal vectors) down to two such signal vectors. Because the
vectors from shifter 202 may be increased in numerical significance
by 18 bit positions, the "width" of compressor 210 needs to be
increased to 57 bits. Hence compressor 210 is shown having 57 bit
positions (i.e., [57:1]).
[0054] The sum and carry vectors output by compressor 210 (each
vector possibly including as many as 57 bits) are applied to muxes
220a and 220b, and also to controllable shifter circuit 230. Mux
circuitry 220a can select any one of various 38-bit subsets of the
57-bit signal vectors output by compressor 210 for application to
an adjacent DSP block to the left as viewed in FIG. 4. (Throughout
this discussion, the bits in all such 38-bit data values or subsets
are of consecutive numerical significance. For example, they may be
relatively low-order bits from the 57-bit source; or they may be
relatively high-order bits from that source. But in either case,
they all have successive bit significance. Selecting such a subset
is typically the result of which one of several groups of
electrical leads is selected as the source of the signals in the
selected subset.) Similarly, mux circuitry 220b can select any one
of various 38-bit subsets of the 57-bit vectors output by
compressor 210 for application to an adjacent DSP block to the
right as viewed in FIG. 4. Muxes 220a and 220b may also have the
ability to output data that is all zeros, if desired. This is a
capability that any of the muxes (e.g., 220, 232, 242, 248, etc.)
and any of the controllable shifters (e.g., 202, 230, etc.) shown
and described throughout this disclosure may have. Some muxes
(e.g., 242, 252, and 262) are depicted with this capability
explicitly shown (e.g., the connection to ground 244 of one set of
the selectable inputs to these muxes). This circuitry is not
repeated for the depiction of all other elements (e.g., 202, 220,
230, 232, 248, etc.) that may include it (to avoid
over-complicating the drawings), nor is this capability expressly
mentioned again in the discussion of all of these other elements,
but it will be understood that all elements of these general kinds
may have this capability.
[0055] Controllable shifter 230 can shift bits applied to it by 18
bits to the left (thereby increasing the numerical significance of
those bits), or by 18 bits to the right (thereby decreasing the
numerical significance of those bits). As a third alternative,
shifter 230 may apply no shift to the data applied to it. All of
elements 220 and 230 may be controlled by programmable fuse or
memory circuit elements like 130 in FIG. 2 (or may be mask
programmable) to select which of their various possible functions
to perform.
[0056] Note that in order to exit the representative DSP block 100
shown in FIGS. 3 and 4, the outputs of muxes 220b must be selected
by redundancy muxes 110a and 110b (which are specific instances of
redundancy muxes 110 shown more generally or generically in FIGS. 1
and 2). The alternative inputs to muxes 110a and 110b are shown by
dotted lines and only in part in FIG. 3 and not at all in FIG. 4
(to avoid over-complicating FIGS. 3 and 4). But it will be apparent
from FIG. 1 that these dotted line inputs come from the muxes 220b
in the DSP block 100 below the FIG. 3 DSP block (or to the left of
the FIG. 4 DSP block).
[0057] Muxes 232 receive 38-bit sum and carry vectors from
redundancy muxes 120a and 120b. Again, redundancy muxes 120a and
120b are specific instances of the type of redundancy muxes that
are shown more generally or generically at 120 in FIG. 1. The solid
line inputs to muxes 120a and 120b come from the muxes 220a in the
DSP block 100 immediately above the FIG. 3 DSP block (or
immediately to the right of the FIG. 4 DSP block). The alternative
dotted line inputs to muxes 120a and 120b (shown only in part in
FIG. 3 and not at all in FIG. 4 to avoid over-complicating FIGS. 3
and 4) come from the muxes 220a in the DSP block 100 two above the
FIG. 3 DSP block (or two to the right of the FIG. 4 DSP block).
Muxes 232 also receive 38-bit sum and carry signal vectors from
leads that come from the redundancy muxes 110a and 110b in the DSP
block 100 immediately below the FIG. 3 DSP block (or immediately to
the left of the FIG. 4 DSP block). Muxes 232 can select either the
vectors from muxes 120a and 120b or the other vectors that muxes
232 receive. Control of muxes 232 can be similar to control of
other variable elements like muxes 220.
[0058] The next element in the representative DSP block 100 shown
in FIGS. 3 and 4 is 4-2 compressor circuit 240. Although different
in size, compressor 240 can be similar to compressor 210. In other
words, compressor 240 can combine the four 38-bit sum and carry
signal vectors it receives into two, further, 38-bit sum and carry
signal vectors. In the event that larger multiplications are being
performed, compressor 240 may need to operate in chain-like
conjunction with the similar compressor 240 in another adjacent DSP
block 100. Accordingly, compressor 240 can additionally receive
lower-order (i.e., lower-numerical-significance) signal bits from
muxes 242. Muxes 242 can select these lower-order bits from a
source of ground potential 244 in the event that there is no
chaining-in from a real lower-order source. Alternatively, if there
is such chaining-in, then muxes 242 get their outputs from
redundancy muxes 120c. Muxes 120c are another instance of muxes
like 120a and 120b. The solid line inputs to muxes 120c are the two
outputs from compressor 240 in the DSP block 100 immediately above
the FIG. 3 block (or immediately to the right of the FIG. 4 block).
The dotted line inputs to muxes 120c (shown only in part in FIG. 3
and omitted entirely from FIG. 4 to avoid over-complication of the
depictions) are the two outputs from the compressor 240 in the DSP
block 100 two above the FIG. 3 block (or two to the right of the
FIG. 4 block). The two outputs that come off the bottom (FIG. 3) or
left (FIG. 4) of compressor 240 are the most significant, highest
order, or overflow output signal bits that result from the
operations performed by compressor 240. Again, these overflow bits
can be chained to the lowest order inputs of an adjacent DSP
block's compressor 240 via muxes 120c and 242 in that adjacent
block if larger multiplications are being performed. Muxes 232 are
controlled to output 0 data signals in the event that there is no
chaining of the DSP blocks.
[0059] The 38-bit sum and carry vectors output by compressor 240
are applied to three-to-two ("3-2") compressor circuit 250. The
final product output by an adjacent DSP block 100 can also be
applied to 3-2 compressor 250 via redundancy muxes 120d and 120e
and muxes 248. Again, this inter-DSP-block routing feature can be
used when certain more complex DSP operations are to be performed.
The arrangement and use of muxes 120d and 120e are similar to the
arrangement and use of other such muxes like 120a and 120b. Thus
the solid line inputs to muxes 120d and 120e come from the primary
outputs of the carry-propagate adder (CPA) circuit 260 in the DSP
block 100 above (FIG. 3) or to the right (FIG. 4) of the block
shown in FIGS. 3 and 4. The alternative dotted line inputs to
redundancy muxes 120d and 120e (again shown only in part in only
FIG. 3) come from the primary outputs of the CPA 260 in the DSP
block 100 two above (FIG. 3) or two to the right (FIG. 4) of the
FIGS. 3 and 4 block. (Muxes 248 are controlled to output 0 data
signals in the event that there is no chaining of the DSP blocks.)
3-2 compressor 250 can combine the sum and carry signal vectors
from compressor 240 with the data output by muxes 248 to produce
further sum and carry signal vectors that are applied to final
carry-propagate adder ("CPA") 260. In the event that the
representative DSP block 100 is involved in a longer-than-38-bit
arithmetic operation, lower-order signal bits can be applied to
compressor 250 and CPA 260 from an adjacent DSP block 100 via muxes
120f and 252 in the case of compressor 250 and via muxes 120g and
262 in the case of CPA 260. Similarly, higher-order signal bits can
be output from elements 250 and 260 to the elements 120f/252/250
and 120g/262/260 in an adjacent DSP block 100 if needed for
chaining multiple elements 250 and multiple elements 260 for longer
arithmetic operations. The arrangement and use of elements 120f,
252, and 244 can be similar to the above-described arrangement and
use of elements 120c, 242, and 244. The same is true for elements
120g, 262, and 244. Accordingly, further explanation of elements
120f/252/244 and 120g/262/244 should not be needed.
[0060] In addition to being substantially directly routable back to
muxes 248 in another adjacent DSP block 100, the final,
up-to-38-bit signal product output by CPA 260 is typically applied
to the other circuitry 20 in the row RN that includes
representative DSP block 100 as shown in FIG. 1.
[0061] We turn now to considering examples of the various operating
modes that one or more DSP blocks 100 can support.
A. 18.times.18 and 18.times.18 Complex
[0062] Relatively simple modes like 18.times.18 multiplication and
18.times.18 complex multiplication (e.g., (a+bi).times.(c+di)) can
be done within one DSP block 100. For example, either multiplier
200 can be used to form sum and carry signal vectors from an
18.times.18 multiplication, and those vectors can be passed down
through subsequent components to CPA 260, which forms the final
product of the multiplication.
B. Sum of Four 18.times.18
[0063] The sum of the products of four 18.times.18 multiplications
can be formed using two adjacent DSP blocks. ("Adjacent" means
immediately adjacent if there is no intervening row that has been
taken out of service by operation of the redundancy circuitry; or,
alternatively, adjacent but for such an intervening out-of-service
row.) In discussions such as the following, each letter such as A,
B, C, etc., denotes or represents an 18-bit input, and each letter
pair like AB, CD, etc., is the multiplication result for a
multiplication of the two 18-bit inputs identified by the letters
in that pair. Such a letter pair may represent either intermediate
sum and carry vectors for such a multiplication result, or the
final multiplication product. Thus AB+CD+EF+GH is the sum of four
products of eight, paired, 18-bit inputs.
[0064] Multiplier 200a in a first DSP block 100 may be used to form
AB. Multiplier 200b in that DSP block may be used to form CD.
Compressor 210 in the first DSP block forms AB+CD. Muxes 220b can
be used to route AB+CD to muxes 232 and compressor 240 in the
adjacent DSP block 100. Multiplier 200a in the adjacent DSP block
can be used to form EF. Multiplier 200b in that adjacent block can
be used to form GH. Compressor 210 in that adjacent block forms
EF+GH. Compressor 240 in that adjacent block forms AB+CD+EF+GH,
which CPA 260 in that same block outputs in final product form.
Note that the block forming AB+CD can be a block to either side of
the block forming EF+GH and AB+CD+EF+GH.
C. Sum of Two 36.times.18
[0065] In discussions such as the following, letters like A, B, C,
etc., have the same meaning as before; letter pairs like AC, BC,
etc., have the same meaning as before; letter pairs like (A,B)
denote 36-bit data words in which the letter on the left refers to
the 18 more significant bits, and the letter on the right refers to
the 18 less significant bits; and letter combinations like (A,B)*C
denote the multiplication result for a multiplication of a 36-bit
data word (e.g., (A,B)) multiplied by an 18-bit data word (e.g.,
C). Thus the objective of the mode being discussed here is to
produce (A,B)*C+(D,E)*F.
[0066] FIG. 5 shows (arithmetically) how the above-mentioned result
is built up. Part 1 of FIG. 5 shows one of the 36.times.18
multiplications to be performed (i.e., (A,B)*C). Part 2 shows the
other of the 36.times.18 multiplications to be performed (i.e.,
(D,E)*F). Part 3 shows how various partial products are produced
and then aligned vertically for addition to produce
(A,B)*C+(D,E)*F. In particular, two partial products of (A,B)*C are
produced as AC and BC, with AC being shifted to the left 18 bit
positions relative to BC. Similarly, two partial products of
(D,E)*F are produced as DF and EF, with DF being shifted 18 bit
positions to the left relative to EF. All of BC, AC, EF, and DF are
then added as shown in Part 3 of FIG. 5.
[0067] FIG. 6 shows how two adjacent DSP blocks 100 can be
configured to perform the above operations. FIG. 6 is simplified by
omitting depiction of the redundancy circuitry and by assuming that
two immediately adjacent DSP blocks 100a and 100b are used. (It
will be understood that all of the previously depicted and
described redundancy circuitry is in fact present in the FIG. 6
circuitry, and that this redundancy circuitry can be used to
effectively put together two DSP blocks that are separated from one
another by an intervening DSP block that cannot be used.) As shown
in FIG. 6, multiplier 200a in DSP block 100a forms AC, multiplier
200b in DSP block 100a forms DF, multiplier 200a in DSP block 100b
forms BC, and multiplier 200b in DSP block 100b forms EF.
Compressor 210 in block 100a compresses (adds) AC and DF.
Compressor 210 in block 100b compresses (adds) BC and EF. The less
significant 18 bits of the output of compressor 210 in block 100a
are routed via elements 220b (block 100a) and 232 (block 100b) to
the compressor 240 in block 100b. (The path of this routing is
highlighted by dotted line 211 in FIG. 6.) This enables block 100b
to form the portion of the final sum to the right of the dotted
line in Part 3 of FIG. 5. The data to the left of the dotted line
in Part 3 of FIG. 5 remains in block 100a and is output by that
block as that part of the final sum.
[0068] More particularly, in block 100a, shifter 230 is used to
shift the outputs of compressor 210 18 bit positions to the right
so that the more significant outputs of compressor 210 that are not
transferred to block 100b are shifted to the less significant
portion of compressor 240 in block 100a. In block 100b shifter 230
shifts the data from compressor 210 18 bit positions to the left,
and the data passing through mux circuitry 232 in block 100b is
applied to the more significant inputs to compressor 240 that are
associated with that mux circuitry 232. The compressor 240 in block
100b is thus able to begin finishing the addition shown to the
right of the dotted line in Part 3 of FIG. 5. Any carry out or
overflow from the most significant end of compressor 240 in block
100b is applied by muxes 242 in block 100a to the least significant
end of compressor 240 in block 100a. This enables both compressors
240 to effectively operate together as one much longer
compressor.
[0069] The data output by each of compressors 240 flows through the
respectively associated compressor 250 to the respectively
associated carry-propagate adder ("CPA") 260, where the final
sum-out data for (A,B)*C+(D,E)*F is produced. Again, any carry
overflow from the most significant end of CPA 260 in block 100b is
applied (as a carry in) to the least significant end of CPA 260 in
block 100a via mux 262 in block 100a. This again enables both of
CPAs 260 to effectively operate together as one much longer CPA.
Thus block 100a outputs the 18 more significant bits of final
product (A,B)*C+(D,E)*F, while block 100b outputs the 36 less
significant bits of that final product.
D. 36.times.36 Mode Using Two DSP Blocks
[0070] In discussions such as the following, letter pairs like
(A,B) again denote a 36-bit data word in which the letter on the
left refers to the 18 more significant bits, and the letter on the
right refers to the 18 less significant bits. Letter pairs like BD
also have basically the same meaning as before (although now such a
letter pair typically refers to the product of two 18-bit portions
of two different 36-bit data words). Thus the objective of the work
being discussed in this section is to produce the product of (A,B)
and (C,D), or (A,B)*(C,D).
[0071] FIG. 7 shows arithmetically how the above-mentioned result
is built up. Part 1 of FIG. 7 shows the multiplication to be
performed. Part 2 shows four partial products that are formed, and
how these four partial products are positioned, left to right,
relative to one another to adjust their relative arithmetic
significances so that they can be added vertically. In particular,
the topmost partial product BD is the product of the 18 less
significant bits in each of multiplicand (A,B) and multiplier
(C,D). The next partial product AD is the product of the 18 more
significant bits in the multiplicand (A,B) and the 18 less
significant bits in the multiplier (C,D). Partial product CB is the
product of the 18 less significant bits in multiplicand (A,B) and
the 18 more significant bits in multiplier (C,D). Partial product
AC is the product of the 18 more significant bits in each of the
multiplicand and the multiplier. As has already been mentioned,
Part 2 of FIG. 7 also shows how partial products AD and CB are
shifted 18 bit positions to the left relative to partial product
BD, as well as how partial product AC is shifted 18 more bit
positions to the left relative to partial products AD and CB. After
such shifting, the four partial products shown in Part 2 of FIG. 7
can be added vertically to produce the desired final product of
(A,B) and (C,D), i.e., (A,B)*(C,D).
[0072] Two adjacent DSP blocks 100 can be used to perform the
arithmetic functions illustrated by FIG. 7. (Again "adjacent" means
either immediately adjacent if no intervening row has been taken
out of service, or adjacent on opposite sides of an intervening row
that has been taken out of service.) FIG. 8 shows how this can be
done using two adjacent DSP blocks 100a and 100b (and again
omitting depiction of the redundancy multiplexers and other
redundancy connections to avoid unduly complicating the FIG.).
[0073] As shown in FIG. 8, multiplier 200a in DSP block 100a forms
partial product AC. Multiplier 200b in DSP block 100a forms partial
product CB. Multipliers 200a and 200b in DSP block 100b form
partial products AD and BD, respectively. Shifter 202 in DSP block
100a shifts partial product AC 18 bit positions to the left
relative to partial product CB, and compressor 210 in that DSP
block adds those two partial products as thus shifted relative to
one another. Shifter 202 in DSP block 100b shifts partial product
AD 18 bit positions to the left relative to partial product BD, and
compressor 210 in that DSP blocks adds those two partial products
as thus shifted relative to one another.
[0074] The 18 least significant bits output by the compressor 210
in DSP block 100a are routed via elements 220b in block 100a and
232 in block 100b to compressor 240 in block 100b, where these bits
are routed to the more significant end of the associated compressor
240 inputs. (Dotted line 213 highlights this routing.) The 18 most
significant bits output by compressor 210 in DSP block 100b are
routed via elements 220a in block 100b and 232 in block 100a to
compressor 240 in block 100a, where these bits are routed to the
less significant end of the associated compressor inputs. (Dotted
line 215 highlights this routing.) Shifter 230 in DSP block 100a
shifts the data output by compressor 210 in that block 18 bits to
the right to prevent the less significant bits that have been
transferred from block 100a to block 100b from also being applied
to compressor 240 in block 100a. Shifter 230 in DSP block 100b
shifts the data output by compressor 210 in that block 18 bits to
the left in order to prevent the more significant bits that have
been transferred from block 100b to block 100a from also being
applied to compressor 240 in block 100b.
[0075] Compressors 240 in DSP blocks 100a and 100b work together to
add the partial product information applied to them (with element
242 in block 100a applying any overflow from the highest-order
(most-significant) bit position in compressor 240 in block 100b to
the lowest-order (least-significant) bit position in compressor 240
in block 100a). Compressor 240 in block 100a is thus beginning to
form the result of addition of data to the left of the vertical
dotted line in Part 2 of FIG. 7, while compressor 240 in block 100b
is performing similarly for the data to the right of that dotted
line.
[0076] The data output by each of compressors 240 flows through the
respective compressor 250 to the respective carry-propagate adder
("CPA") 260. Again, element 262 in block 100a applies any carry out
from the most significant end of CPA 260 in block 100b to the least
significant end of CPA 260 in block 100a. The final outputs of
these two CPAs 260 collectively comprise the final product
(A,B)*(C,D), with the outputs of CPA 260 in block 100a constituting
the more significant bits of that final product, and with the
outputs of CPA 260 in block 100b constituting the less significant
bits of that final product.
E. 54.times.54 Mode Using Four DSP Blocks
[0077] Another example of how DSP blocks 100 in accordance with
this invention can be used is illustrated by FIGS. 9 and 10. In
this discussion letter triplets like (A,B,C) refer to 54-bit data
words in which letter A denotes the 18 most-significant bits,
letter B denotes the 18 bits of intermediate arithmetic
significance, and letter C denotes the 18 least-significant bits.
Letter pairs like AF refer to a partial product of 18 bits A from
one 54-bit data word times 18 bits F from another 54-bit data word.
The objective of the mode being discussed in this section is to
produce the product of two 54-bit data words, i.e., the product of
(A,B,C) times (D,E,F), or (A,B,C)*(D,E,F). It is assumed, however,
that exact precision for 108 bits is not required for the product.
Accordingly, the least significant partial product CF is not
computed or used to produce the final (approximate) product.
[0078] FIG. 9 shows arithmetically how the above-mentioned product
is built up from multiple partial products. FIG. 10 shows four DSP
blocks 100a-100d that can be used to build up the product as shown
in FIG. 9. (Again, for simplicity, FIG. 10 assumes that four
immediately adjacent DSP blocks 100 can be used because no DSP
block in this range is out of service. Thus FIG. 10 omits depiction
of the redundancy circuitry shown and described elsewhere in this
specification. But that redundancy circuitry is preferably present
and can be used to enable another adjacent DSP block 100 to be used
to help perform the functions described in connection with FIG. 10
if one of blocks 100a-d must be taken out of service.) Part 1 of
FIG. 9 shows the multiplication to be performed. Part 2 of FIG. 9
shows the partial products that are formed and appropriately
combined in the two left-hand DSP blocks 100a and 100b. Part 3 of
FIG. 9 shows the partial products that are formed and appropriately
combined in the two right-hand DSP blocks 100c and 100d. The 36
more significant bits from Part 3 of FIG. 9 (i.e., the bits to the
left of the dotted line in Part 3) are added to the 36 less
significant bits from Part 2 of FIG. 9 (i.e., the bits to the right
of the dotted line in Part 2), primarily in the final stages of DSP
block 100b. Thus the 72 more significant bits of the product are
output by DSP blocks 100a and 100b. An additional 18 less
significant bits of the product are available from DSP block 100d.
As mentioned earlier, still less significant bits of the product
are not computed and are not available. Thus the "product" in this
case is only approximate, although it is approximate to a very high
degree of precision. In particular, this is due to the fact that
the least significant partial product CF is not computed or used
anywhere in FIGS. 9 and 10.
[0079] Considering FIGS. 9 and 10 now in more detail, it will be
noted that the work done in DSP blocks 100a and 100b (as shown in
Part 2 of FIG. 9) is initially the same as what is shown in FIGS. 7
and 8 and described above in connection with those FIGS. In
particular, this work is use of two DSP blocks 100a and 100b to
form the product of two 36-bit data words (in this case the product
(A,B)*(D,E)). Because this aspect of the functionality of two DSP
blocks has already been fully described in connection with FIGS. 7
and 8, it will not be necessary to describe it in detail again in
connection with Part 2 of FIG. 9 or DSP blocks 100a and 100b in
FIG. 10.
[0080] Turning now to Part 3 of FIG. 9 and DSP blocks 100c and 100d
in FIG. 10, it will be noted that the work done in these two DSP
blocks is initially somewhat similar to what is shown in FIGS. 5
and 6 and described above in connection with those FIGS. However,
the dotted line in Part 3 of FIG. 9 is in a different location than
the dotted line in Part 3 of FIG. 5, so the way in which DSP blocks
100c and 100d in FIG. 10 do this work is somewhat different than
the flow described above for FIG. 6. The somewhat different flow
used in FIG. 10 will be described in the next paragraph.
[0081] Multipliers 200a and 200b in DSP block 100c form partial
products AF and DC, respectively. Compressor 210 in block 100c adds
these two partial products together. Multipliers 200a and 200b in
DSP block 100d form partial products BF and EC, and the compressor
210 in that block adds these two partial products together. Routing
220a in block 100d and 232 in block 100c applies the 18 more
significant bits output by compressor 210 in block 100d to the less
significant end of compressor 240 in block 100c. Shifter 230 in
block 100d shifts the outputs of the compressor 210 in that block
18 bit positions to the left. Compressor 240 in block 100c
compresses the four vectors applied to it down to two vectors,
which flow down through the compressor 250 in that block to the CPA
260 in that block. (This is basically the final addition work
required to the left of the dotted line in Part 3 of FIG. 9.) The
outputs of compressor 240 in block 100d similarly flow down through
the compressor 250 in that block to the CPA 260 in that block.
(This corresponds to what is to the right of the dotted line in
Part 3 of FIG. 9.) The CPAs 260 in blocks 100c and 100d work
together to produce the final sum of the work shown in Part 3 of
FIG. 9. Elements 248 in block 100b are used to route the 38 more
significant bits of that result (output by the CPA 260 in block
100c) into block 100b for addition to the work being done in blocks
100a and 100b (as shown in Part 2 of FIG. 9). Thus the final result
is (1) the 72 more significant bits of (A,B,C)*(D,E,F) being output
by the CPAs 260 in blocks 100a and 100b, and (2) the 18 more (less
significant) bits of that (approximate) product being output by the
CPA 260 in block 100d.
[0082] Recapitulating some aspects of the foregoing, and also in
some respects extending what has been said, a digital signal
processing ("DSP") block (e.g., 100) may include first and second
N-bit (e.g., 18-bit) multiplier circuits (e.g., 200a and 200b). The
DSP block may further include first shifter circuitry (e.g., 202)
for shifting outputs of the first multiplier circuit by a
selectable one of (1) zero bit positions and (2) N bit positions
toward greater arithmetic significance. The DSP block may still
further include first compressor circuitry (e.g., 210) for
additively combining outputs of the first shifter circuitry and the
second multiplier circuit. The DSP block may yet further include
circuitry (e.g., 220a and 220b) for selectively routing outputs of
the first compressor circuitry to first and second other DSP
circuit blocks that are on respective opposite sides of the DSP
circuit block. The DSP block may still further include second
shifter circuitry (e.g., 230) for shifting outputs of the first
compressor circuitry by a selectable one of (1) zero bit positions,
(2) N bit positions toward greater arithmetic significance, and (3)
N bit positions toward lesser arithmetic significance. The DSP
block may yet further include second compressor circuitry (e.g.,
240) for additively combining any outputs received from the first
compressor circuitry in either of the first and second other DSP
circuit blocks.
[0083] In a DSP circuit block as described above, the routing
circuitry (e.g., 220a and/or 220b) may be controllable to select
for routing any one of a plurality of subsets of the outputs of the
first compressor circuitry (e.g., 210). These selectable subsets
may include (1) a subset including a most significant output bit
position of the first compressor circuitry, and (2) a subset
including a least significant output bit position of the first
compressor circuitry.
[0084] In a DSP circuit block as described above the second
compressor circuitry (e.g., 240) may include overflow output
circuitry (e.g., output leads from most significant end of
compressor 240 to adjacent DSP block) for applying overflow output
signals of the second compressor circuitry to the first other DSP
circuit block, and overflow input circuitry (e.g., 242) for
selectively receiving overflow output signals of the second
compressor circuitry in the second other DSP circuit block.
[0085] A DSP circuit block as described above may further include
third compressor circuitry (e.g., 250) for additively combining
outputs of the second compressor circuitry and any further outputs
received from the second other DSP circuit block, and further
routing circuitry (e.g., 248) for selectively routing outputs of
the third compressor circuitry, as further outputs, to the first
other DSP circuit block. In such a DSP circuit block, the third
compressor circuitry (e.g., 250) may comprise overflow output
circuitry (e.g., output leads from most significant end of
compressor 250 to adjacent DSP block) for applying overflow output
signals of the third compressor circuitry to the first other DSP
circuit block, and overflow input circuitry (e.g., 252) for
selectively receiving overflow output signals of the third
compressor circuitry in the second other DSP circuit block.
[0086] In a DSP circuit that includes the above-mentioned further
routing circuitry (e.g., 248), that further routing circuitry may
include carry-propagate adder ("CPA") circuitry (e.g., 260) for
operating on the outputs of the third compressor circuitry (e.g.,
250) to produce the further outputs. The CPA circuitry may include
carry-out circuitry (e.g., output lead from most significant end of
CPA 260 to adjacent DSP block) for applying a carry out signal of
the CPA circuitry to the first other DSP circuit block, and
carry-in circuitry (e.g., 262) for selectively receiving a carry
out signal of the CPA circuitry in the second other DSP circuit
block.
[0087] A DSP circuit block as described above may further include
redundancy circuitry (e.g., 110 and/or 120) for allowing the first
other DSP circuit block to be a selectable one of (1) another DSP
circuit block that is immediately adjacent to the DSP circuit
block, and (2) yet another DSP circuit block that is not
immediately adjacent to the DSP circuit block. For example, if the
DSP circuit block is in row R7 in FIG. 1, the first other DSP block
can be either (1) in row R8 or (2) in row R9. In such an
arrangement the yet another DSP circuit block (e.g., the one in row
R9) is immediately adjacent to the another DSP circuit block (e.g.,
the one in row R8).
[0088] A DSP circuit block as described above may also include
further redundancy circuitry (e.g., 110 and/or 120) for allowing
the second other DSP circuit block to be a selectable one of (1)
still another DSP circuit block that is immediately adjacent to the
DSP circuit block, and (2) still a further other DSP circuit block
that is not immediately adjacent to the DSP circuit block. For
example, if the DSP circuit block is in row R7 in FIG. 1, the
second other DSP block can be either (1) in row R6 or (2) in row
R5. In such an arrangement the still a further other DSP circuit
block (e.g., the one in row R5) is immediately adjacent to the
still another DSP circuit block (e.g., the one in row R6).
[0089] In accordance with other possible aspects of the invention,
DSP circuitry (e.g., 10) may comprise a plurality of DSP circuit
blocks (e.g., 100), each of which is capable of performing DSP
operations (e.g., 200, 202, 210, 230, 240, etc.) on signals applied
to that circuit block (e.g., A, B, C, etc.), each of the DSP
circuit blocks may include circuitry (e.g., 220, 232, 242, etc.)
for selectively routing outputs of at least some of the DSP
operations to first and second other ones of the DSP circuit blocks
that are on respective opposite sides of the DSP circuit block. In
such cases, the circuitry for selectively routing may include
redundancy circuitry (e.g., 110 and/or 120) for allowing the first
other DSP circuit block for a DSP circuit block to be a selectable
one of (1) another DSP circuit block that is immediately adjacent
to that DSP circuit block, and (2) yet another DSP circuit block
that is not immediately adjacent to that DSP circuit block. For
example, if the DSP circuit block is in row R7 in FIG. 1, the first
other DSP block can be either (1) in row R8 or (2) in row R9. In
such an arrangement the yet another DSP circuit block (e.g., the
one in row R9) is immediately adjacent to the another DSP circuit
block (e.g., the one in row R8).
[0090] The circuitry for selectively routing in DSP circuitry as
described above may further include further redundancy circuitry
(e.g., 110 and/or 120) for allowing the second other DSP circuit
block for a DSP circuit block to be a selectable one of (1) still
another DSP circuit block that is immediately adjacent to that DSP
circuit block, and (2) still a further other DSP circuit block that
is not immediately adjacent to that DSP circuit block. For example,
if the DSP circuit block is in row R7 in FIG. 1, the second other
DSP block can be either (1) in row R6 or (2) in row R5. In such an
arrangement the still further other DSP circuit block (e.g., the
one in row R5) is immediately adjacent to the still another DSP
circuit block (e.g., the one in row R6).
[0091] In DSP circuitry as described above, the circuitry for
selectively routing of each of the DSP circuit blocks may
selectively route the outputs to inputs of the first and second
other ones of the DSP circuit blocks that are downstream from some
but not all DSP operations that the first and second other ones of
the DSP circuit blocks are capable of performing. For example,
selective routing circuitry 220 can route outputs of compressor 210
in one DSP block to inputs of a compressor 240 in another DSP
block, and compressor 240 is downstream from some (but not all)
other DSP operations in the other DSP block (e.g., it is downstream
from operations 200, 202, 210, and 230, but it is upstream from
operations 250 and 260).
[0092] In accordance with still other possible aspects of the
invention, DSP circuitry (e.g., 10) may include a plurality of DSP
circuit blocks (e.g., 100), each of which is capable of performing
a plurality of DSP operations (e.g., 200, 202, 210, 230, 240, etc.)
one after another in succession. Each of the DSP circuit blocks may
further include circuitry (e.g., 220) for selectively routing
outputs of at least one of the DSP operations (e.g., 210) of that
DSP circuit block to first and second other ones of the DSP circuit
block that are on respective opposite sides of that DSP circuit
block. The circuitry for selectively routing in each DSP circuit
block may selectively route at least some of the outputs to inputs
to DSP operations in the first and second other DSP circuit blocks
that are intermediate in the succession of DSP operations in those
other DSP circuit blocks. For example, routing circuitry 220 can
route outputs of a DSP block to inputs of compressor 240 in another
DSP block, and compressor 240 is intermediate in the succession of
DSP operations (i.e., it is preceded by DSP operations like 200 and
202, and it is followed by DSP operations like 250 and 260).
[0093] In DSP circuitry as described above, at least some of the
inputs may be inputs to DSP operations at a different point in the
succession in the DSP blocks than the point in the succession in
the DSP circuit block from which the circuitry for selectively
routing received the outputs selectively routed to those at least
some inputs. For example, compressor 210, from which routing
elements 220 get outputs to apply to compressor 240 in another DSP
block, is at a different point in the succession of DSP operations
in the first-mentioned DSP block than the point at which compressor
240 is in the succession of DSP operations in the second-mentioned
DSP block.
[0094] In DSP circuitry as described above, each DSP circuit block
may further include second circuitry (e.g., the lead from the most
significant end of compressor 240 to an adjacent DSP block) for
selectively routing a second output of a DSP operation in that DSP
circuit block to an input of a same DSP operation in the first
other DSP circuit block. For example, the immediately
above-mentioned lead and element 242 allows overflow signals to go
from the compressor 240 in one DSP block to the compressor 240 in
another DSP block.
[0095] In DSP circuitry as described above, each DSP circuit block
may further include third circuitry (e.g., 248) for selectively
routing outputs of a final DSP operation (e.g., 260) in that DSP
circuit block to inputs to a DSP operation (e.g., 250) in the first
other DSP circuit that is at an intermediate point in the
succession in that other DSP circuit block.
[0096] DSP circuitry as described above may further include
redundancy circuitry (e.g., 110 and/or 120) for allowing the first
other DSP circuit block of each of the DSP circuit blocks to be a
selectable one of (1) another of the DSP circuit blocks that is
immediately adjacent to the DSP circuit block, and (2) yet another
of the DSP circuit blocks that is not immediately adjacent to the
DSP circuit block. For example, if a DSP block 100 is in row R7 in
FIG. 1, the another DSP block may be in row R8 and the yet another
DSP block may be in row R9. In addition, the yet another DSP block
(e.g., the one in row R9) may be immediately adjacent to the
another DSP block (e.g., the one in row R8).
[0097] DSP circuitry as described above may also include further
redundancy circuitry (e.g., 110 and/or 120) for allowing the second
other DSP circuit block of each of the DSP circuit blocks to be a
selectable one of (1) still another of the DSP circuit blocks that
is immediately adjacent to the DSP circuit block, and (2) still a
further another one of the DSP circuit blocks that is not
immediately adjacent to the DSP circuit block. For example, if the
DSP circuit block is in row R7 in FIG. 1, the still another DSP
block may be in row R6, and the still a further another DSP block
may be in row R5. The still a further another one of the DSP
circuit blocks (e.g., the one in row R5) may be immediately
adjacent to the still another of the DSP circuit blocks (e.g., the
one in row R6).
[0098] The circuitry of this invention allows summations to be
performed by bidirectional shifting between DSP circuit blocks. For
example, a value can be shifted from a first DSP block to a second
DSP block to the right of the first block and combined (e.g.,
compressed) with another signal in the second block. The result of
this combination can then be shifted back to the left (i.e., to the
first block) and combined with other signals in the first block.
This shifting back can occur (for example) through (1) the carry
vectors from compressor to compressor (compressors 240 or
compressors 250), (2) the carry bit of the CPA 260, or (3) the
output of the CPA 260 to the 3-2 compressor 250.
[0099] Any of the multiplexers employed in the DSP circuitry of
this invention can be of the type that can selectively (i.e.,
controllably) output zero (0) data. This also includes any of the
controllable shifters employed herein.
[0100] Another illustrative embodiment of a DSP circuit block 500
in accordance with certain possible aspects of this disclosure is
shown in FIG. 11. The inputs to DSP block 500 can be dynamic (e.g.,
electrical signals representing digital (binary) numerical values
that are time-varying during normal (user mode) operation of the
integrated circuit (e.g., 10) that includes the DSP block).
Alternatively, some of the inputs to DSP block 500 can be static
(e.g., electrical signals representing digital (binary) numerical
values that are constant or relatively constant over time during
normal operation of the integrated circuit). For example, inputs
ax, ay, bx, and by may be dynamic inputs; inputs c and d may be
static inputs. Static inputs may come from memory (e.g., so-called
configuration random access memory or CRAM) on the IC that is
programmed during an initial set-up (configuration or
initialization) phase of operation of the IC (prior to subsequent
normal or user mode operation of the IC). (An illustration of a
static signal source is element 130 in FIG. 2. Another illustration
of such a source is element 602 in FIG. 12.)
[0101] In the illustrative embodiment shown in FIG. 11, each of
inputs ax, bx, ay, by, c, and d comprises a set (plurality) of 18
signals supplied (applied) in parallel. Although 18 is presently
thought to be a highly desirable number, it will be understood that
this number can be larger or smaller, if desired. This basic input
bus width will be sometimes referred to as N in subsequent
discussion herein. Thus N is 18 in the embodiment shown in FIG. 11
and similarly sized embodiments. But it will be understood that N
can be larger or smaller than 18 in other embodiments in accordance
with this disclosure.
[0102] In the illustrative embodiment shown in FIG. 11, dynamic
inputs have the option to be registered by input registers 501
(e.g., flip-flop circuits clocked by a clock signal). See also FIG.
12, which shows how typical inputs ay can be applied to typical
registers 501ay, and also to one selectable set of inputs to
multiplexer ("mux") circuitry 601. Registers 501ay register the ay
signals applied to them in response to and in synchronism with an
applied clock signal. After thus being registered, the registered
ay signals are applied to a second selectable set of inputs to mux
601. Mux 601 can select either of its selectable sets of inputs to
be its output signals, depending on the logical state (e.g., binary
1 or binary 0) of the selection control input signal applied to it
from circuit element 602. For example, element 602 may be a CRAM
element (similar, e.g., to elements like element 130 in FIG. 2).
FIG. 12 thus shows how the outputs of mux 601 can be either
registered or unregistered signals ay. FIG. 12 is illustrative of
how all of registers 501 may be augmented with circuitry that
allows the registers to be either used or bypassed.
[0103] Certain inputs can be routed via an input cascade chain, if
desired. For example, one set of inputs can be cascaded in from a
similar DSP block 500 to the right of the depicted DSP block. Such
cascade in inputs go to one selectable set of inputs to mux 515bx
(see also FIG. 13). Depicted inputs bx are applied to the other
selectable set of inputs to mux 515bx. Mux 515bx is controlled by a
selection control input signal (e.g., from CRAM 610) to select
either of its selectable input sets to be its outputs (depending on
the logical state of the selection control input signal). This
cascade chain continues through (selectively bypassable) registers
501bx, mux 515ax, (selectively bypassable) registers 501ax, to the
cascade out leads, which go the next adjacent DSP block 500 to the
left of the depicted DSP block. The depicted cascade out leads
become the cascade in leads in the DSP block to the left. Thus all
of the x inputs (i.e., ax and bx) from each module or block 500
form a cascade chain.
[0104] An example of an application of the circuitry in which the
immediately above-described cascade chain may be used is
implementation of a finite-impulse-response ("FIR") digital filter.
Such filters are discussed in more detail later is this
specification. Here, however, it will be sufficient to note that
FIG. 26 shows one form of such a filter in which successive samples
x[n] of an input signal are successively delayed by one sample time
period in each of successive sample delay elements 820-1, 820-2,
etc. In each sample time period, each sample or delayed sample is
multiplied by a respective filter coefficient c1, c2, etc., in a
respective one of multipliers 810-1, 810-2, etc. The immediately
above-described cascade chain can be used to provide the input
sample delay chain shown along the bottom line in FIG. 26.
Multipliers 505 and 506 in one or more DSP blocks 500 can provide
the multiplications (by coefficients c1, c2, etc.) of the variously
delayed input samples as shown across the middle of FIG. 26.
Circuitry downstream from the multipliers 505 and 506 in these DSP
blocks can provide some or all of the addition of multiplication
products shown along the top line in FIG. 26. Again, additional
discussion of FIR filters will be found later in this
specification.
[0105] Circuitry 502 is pre-adder circuitry. It can add two of the
dynamic inputs (e.g., ax and ay, or bx and by) together. The
results of the pre-adder then go to controllable routing circuitry
503 to become two of several sets of possible inputs to the
downstream multipliers (described later). Pre-adder 502 may provide
a more cost-efficient and faster alternative to having the
above-described addition done outside of DSP block 500. The eight
possible sets of inputs to routing circuitry 503 are ax, bx, ay,
by, c, d, (ax+ay), and (bx+by).
[0106] Routing circuitry 503 (see also FIG. 14) is controllable
(e.g., by signals from CRAM bits 620) to route signals from any of
its sets of inputs to various ones of its output signal groups. In
particular, the input-to-output routing performed by circuitry 503
is selected (chosen, controlled) by the logical states of the
control signals that are applied to circuitry 503 (e.g., from CRAM
bits 620). As just one example of such possible signal routings,
circuitry 503 may route its c inputs to one of the groups of its
outputs that ultimately go to multiplier circuitry 505, and its by
inputs (which may have been registered by registers 501by) to the
other of the groups of its outputs that ultimately go to multiplier
circuitry 505. At the same time as the above routing is being
provided, circuitry 503 may route its d inputs to one of its groups
of outputs that go to multiplier circuitry 506, while routing its
ay inputs (which may have been registered by register 501ay) to the
other of its groups of outputs that go to multiplier 506. Again,
the foregoing is only one example of many possible input-to-output
signal routing patterns that circuitry 503 may provide. Other
possible routings will be apparent from the various ways in which
multipliers 505 and 506 are described as being constructed and used
in subsequent paragraphs in this disclosure. In particular, routing
circuitry 503 can route any input signal to any multiplier input
that needs to receive that input signal in order to accomplish any
multiplication function that is described below. For example, to
enable multiplier 505 to perform its part of the below-described
27.times.27 multiplication of 27 "x" bits by 27 "y" bits, circuitry
503 can route 27 bits from the 36 ax and bx inputs and 27 bits from
the 36 ay and by inputs to multiplier 505. At the same time, to
enable multiplier 506 to perform its part of this 27.times.27
multiplication, circuitry 503 can route the 18 more significant
bits of the above-mentioned 27 "x" bits and the 18 more significant
bits of the above-mentioned "y" bits to multiplier 506.
[0107] Although depicted like a single multiplexer ("mux") in FIG.
11, circuitry 503 may actually be an array or network of several
multiplexers ("muxes"). For example, FIG. 15 shows that circuitry
503 may include a separate mux 630r, 630s, 630t, etc., for each
group of outputs that it is desired to be able to select
independently or at least partly independently. Each such mux has
as its selectable inputs all the various input signal groups (or
subgroups) from which that mux may need to select the signals to be
output. Each mux 630 is controlled (e.g., by an associated subset
of CRAM bits 620 (such as subset 620r controlling mux 630r)) to
select the desired group of its inputs to be its outputs. If some
group of inputs will never be needed as a particular group of
outputs from circuitry 503, that input group can be omitted from
the input groups available to the mux 630 that selects and provides
that group of output. Similarly, some output groups shown in FIG.
11 may actually be composites of signals from two input groups
(see, for example, the 27-bit output groups mentioned in the
immediately preceding paragraph). Such composite output groups may
be most conveniently provided by using two muxes 630 (e.g., one mux
630 providing the 18-bit portion of such a composite output group,
and another mux 630 providing the 9-bit part of such a composite
output group). Muxes 630 that need to handle only subsets of basic
N-bit groups (e.g., the 9-bit groups that were just mentioned) have
input bit and output bit group sizes that are smaller than N (e.g.,
9-bit or 0.5N input and output groups). Moreover, each of these
smaller input bit groups may come from predetermined portions of
larger (N-bit) input groups. For example, such a predetermined
portion may be the 9 less-significant bits of 18-bit input group
ay, or the 9 more-significant bits of 18-bit input group ay. Thus,
whereas FIG. 15 shows representative muxes 630r and 630s each
receiving and outputting groups of 18 signals, FIG. 15 shows
representative mux 630t receiving and outputting smaller groups of
9 signals.
[0108] From the foregoing it will be seen that circuitry 503 is
preferably highly flexible routing circuitry that can connect any
of its various, selectable, input signal groups (or subsets of such
groups) to the various ones of its output signal groups that may
need those inputs to support any of the multiplier functions that
will be described later in this disclosure.
[0109] Returning to FIG. 11, the output signal groups from
circuitry 503 that go to multiplier circuitry 506 are shown as
going directly to that multiplier. However, the output signal
groups from circuitry 503 that go to multiplier circuitry 505 are
shown as optionally passing through systolic delay registers 504.
As shown in more detail in FIG. 16, the inputs to representative
systolic delay registers 504p are also applied to one selectable
set of inputs to multiplexer ("mux") circuitry 640p. The outputs of
systolic registers 504p are applied to the other selectable set of
inputs to mux 640p. A selection control input signal (e.g., from
CRAM 642p) controls which set of its selectable inputs mux 640p
outputs. Systolic registers 504p are clocked by an applied clock
signal so that any signals passing through the systolic registers
are delayed by one clock cycle. Representative mux 640p allows the
signals applied to representative systolic registers 504p to either
pass through those registers (thereby delaying those signals by one
clock cycle) or to bypass those registers (thereby not delaying
those signals). The purposes for which systolic registers 504 may
be used will be described later in this specification.
[0110] The default condition of multipliers 505 and 506 is as two
18.times.18 multipliers. In general (i.e., in all modes of
operation), each of multipliers 505 and 506 produces two product
vectors in redundant form (e.g., one sum vector and one carry
vector as described in more detail earlier for other embodiments).
Thus in 18.times.18 mode, each multiplier 505 and 506 produces a
36-bit sum vector and a 36-bit carry vector. Circuitry downstream
from multipliers 505 and 506 can be used to further process (e.g.,
additively combine) these vectors in various ways to produce one or
more final products (or a final sum of products) as will be fully
described later in this specification. Preliminarily, however, more
will first be said about multiplier 505.
[0111] Multiplier 505 is modified (e.g., as compared to multiplier
506) so that, together, multipliers 505 and 506 can perform a
27.times.27 multiplication. FIGS. 17 and 18 show an illustrative
embodiment for doing this. (Note that FIG. 17 omits depiction of
systolic registers 504 because they are not used (i.e., they are
bypassed) when a DSP block 500 is performing a 27.times.27
multiplication.)
[0112] As shown in FIG. 17, multiplier 505 includes a 9.times.27
multiplier circuitry part 650a and a 9.times.18 multiplier circuit
part 650b. Multiplier 506 has the unmodified 18.times.18
construction. Assume that it is desired to multiply the 27-bit
multiplicand (x3, x2, x1) by the 27-bit multiplier (y3, y2, y1).
(See also the first two lines of FIG. 18.) Each segment of each of
these 27-bit data words (i.e., each of x3, x2, x1, y3, y2, and y1)
has 9 bits. The most arithmetically significant 9-bit segment in
each of these words is the x3 or y3 segment. The least
arithmetically significant 9-bit segment in each of these words is
the x1 or y1 segment.
[0113] Routing circuitry 503 routes segments x3 and x2 to one axis
(e.g., the multiplicand axis or inputs) of multiplier 506. Routing
circuitry 503 routes segments y3 and y2 to the other axis (e.g.,
the multiplier axis or inputs) of multiplier 506. Routing circuitry
503 routes segments x3, x2, and x1 to the multiplicand axis or
inputs of multiplier 650a. Routing circuitry 503 routes segment y1
to the multiplier axis or inputs of multiplier 650a. Routing
circuitry 503 routes segments y3 and y2 to the multiplicand axis or
inputs of multiplier 650b. Routing circuitry 503 routes segment x1
to the multiplier axis or inputs of multiplier 650b. Each
multiplier circuit 506, 650a, and 650b multiplies the data applied
to its multiplicand axis by the data applied to its multiplier
axis. (Of course, it will be understood that the multiplicand and
multiplier in a multiplication are interchangeable, so it does not
matter which data is supplied as, or referred to as, the
multiplicand, or which data is supplied as, or referred to as, the
multiplier. The terms multiplicand and multiplier are therefore
used arbitrarily and interchangeably herein. Reference to some data
as a multiplicand or as a multiplier is not intended to be limiting
in any way, and in other embodiments the opposite terms could be
employed without departing from the scope or spirit of this
disclosure.) Each multiplier 506, 650a, and 650b produces a
redundant form sum output vector and a redundant form carry output
vector, which are collectively indicative of the multiplication
product formed by that multiplier. (Again, this redundant form
(i.e., sum vector and carry vector) product formation is as
described earlier in this specification.)
[0114] At this point it may be helpful to interrupt the discussion
of FIG. 17 in order to consider FIG. 18, which shows the
27.times.27 multiplication being performed by the FIG. 17
circuitry. The first two lines of FIG. 18 show the multiplication
to be performed (i.e., 27-bit multiplicand x3, x2, x1 in line 1,
and 27-bit multiplier y3, y2, y1 in line 2). Line 3 in FIG. 18
shows the result of the 9.times.27 multiplication performed by
multiplier 650a. Line 4 of FIG. 18 shows the result of the
9.times.18 multiplication performed by multiplier 650b. Note that
the partial product that results from the 650b multiplication must
be shifted 9 bits to the left in order for the bits in this partial
product to line up (vertically) with bits having the same
arithmetic significance in the multiplier 650a partial product.
This shift can be accomplished by how the output leads of
multiplier 650a and 650b enter 4:2 compressor circuitry 652
relative to one another. 4:2 compressor 652 (which can be similar
to such compressor circuitry described earlier in this
specification) receives the sum and carry vectors output by
multipliers 650a and 650b, and produces further redundant form sum
and carry vectors indicative of a further partial product that is
the sum of lines 3 and 4 in FIG. 18. These further sum and carry
vectors are the final outputs of multiplier 505.
[0115] The last major partial product that is needed for the
27.times.27 multiplication being considered is the partial product
that results from the 18.times.18 multiplication shown in line 5 of
FIG. 18. This partial product is formed by multiplier 506 in FIG.
17. As FIG. 18 shows, this partial product needs to be shifted 18
bit positions to the left (increased in arithmetic significance)
relative to the partial product shown in line 3 of FIG. 18. This
shift is performed by controllable shifter circuits 507s (for the
redundant form sum vector output by multiplier 506) and 507c (for
the redundant form carry vector output by multiplier 506). CRAM
bit(s) 654 control whether elements 507 change the arithmetic
significance of the data passing through them. As for other similar
circuit elements used throughout this disclosure, a shifter circuit
like 507s or 507c can shift the arithmetic significance of a data
bit applied to it by controllably routing that bit to a shifter
output lead that has a different (greater or lesser) arithmetic
significance (relative position in a bus having several parallel
data leads) than the arithmetic significance of the input lead on
which that data bit was supplied to the shifter circuit. If no
change in arithmetic significance is needed, then the shifter
circuit is controlled to route the input data bit to a shifter
output lead having the same arithmetic significance as the shifter
input lead supplying that data bit. In the case of performing a
27.times.27 multiplication, shifter circuits 507s and 507c are
controlled to shift the outputs of multiplier 506 18 bits to the
left (thereby increasing the arithmetic significance of each bit in
these outputs by 18 bit positions).
[0116] The outputs of multipliers 505 and 506 are applied to 4:2
compressor circuitry 508. This circuitry can again be similar to
other compressor circuitry described earlier in this specification.
Thus circuitry 508 compresses the two sum vectors and the two carry
vectors applied to it down to one further sum vector and one
further carry vector. These further vectors are indicative of the
final product of the 27.times.27 multiplication being performed. In
particular, these further vectors (output by compressor 508) are
indicative of the result shown in line 6 of FIG. 18. That actual
final result can be produced and output from DSP block 500 by the
further elements of the DSP block that are downstream from
compressor 508 in FIG. 11 and that are further described later in
this specification. This shows how one DSP block 500 can perform
one complete 27.times.27 multiplication.
[0117] Note that for the operations shown in FIG. 18 to perform
correctly on "signed" operands (i.e., values like x and/or y that
can be either positive numbers or negative numbers), it may be
necessary for there to be "sign-extension" from the actual data
shown (e.g., in lines 3 and 4) up to the most significant position
of the result (e.g., in line 6). This may mean including 18 more
leading bits of all Os (e.g., for any positive number) or all is
(e.g., for any negative number) to the left of the data shown in
lines 3 and 4 in FIG. 18. As a general matter, sign-extension is a
conventional technique that is well known to those skilled in the
art. Accordingly, any required sign-extension is generally assumed
herein to be present and is not actually shown in FIG. 18 (or any
other FIGS. herein of a generally similar nature) to avoid
unnecessarily complicating the drawings and/or obscuring the
present disclosure. Examples of other FIGS. where sign-extension
may be employed without actually being shown herein are FIGS. 5, 7,
9, 22, and 31. Sign-extension may also be employed for a certain
area in FIG. 20, but that aspect of FIG. 20 is more specifically
discussed later in this specification.
[0118] For completeness, FIG. 19 shows how the multiplier 505
portion of the FIG. 17 circuitry can be used when that multiplier
505 circuitry is only needed to perform an 18.times.18
multiplication (e.g., of x2 and x1 by y2 and y1). In this case, mux
630u in routing circuitry 503 selects 9-bit segments x2 and x1 for
application to the 18-bit axis of multiplier 650b. Mux 630w also
selects x2 and x1 for application to the 18 less significant inputs
to the 27-bit axis of multiplier 650a. Mux 630t selects 9-bit
segment y2 for application to the 9-bit axis of multiplier 650b.
Mux 630x selects 9-bit segment y1 for application to the 9-bit axis
of multiplier 650a. Null data is applied to the 9 more significant
inputs to the 27-bit axis of multiplier 650a. (If the relevant
operand (i.e., operand x) is an "unsigned" number, then the
just-mentioned "null data" may be all binary Os. For a "signed"
operand, the most significant 9 bits of the 27-bit input should be
the "sign-extension" from the lower-order (less significant) 18
bits. Thus (for a signed operand) the above-mentioned "null data"
can be either all is (e.g., for a negative operand) or all Os
(e.g., for a positive operand). For convenience herein, the term
"null," "null data," or the like is used for all of the various
kinds of data referred to in this parenthetical, i.e., either (a)
"sign-extension" data for a signed operand, or (b) leading Os for
an unsigned operand.)
[0119] FIG. 20 shows the operations that result from the data
routing described in the preceding paragraph. In particular, the
multiplication result shown in line 3 of FIG. 20 is produced by
multiplier circuit 650a. (Note again that "null" in line 3 of FIG.
20 may refer to "sign-extension" data from the lower-order 27 bits
in that line.) The multiplication result shown in line 4 of FIG. 20
is produced by multiplier circuit 650b. Compressor 652 produces the
result (albeit in redundant form) shown in line 5 of FIG. 20. In
this way multiplier circuit 505 can perform an 18.times.18
multiplication as an alternative to performing part of a
27.times.27 multiplication.
[0120] FIG. 21 shows the lower part of FIG. 11 in somewhat more
detail. As has already been mentioned in connection with FIG. 17,
circuitry 508 is a 4:2 compressor, which (like other similar 4:2
compressor circuitry throughout this disclosure) combines two
applied redundant form sum vectors and two applied redundant form
carry vectors to produce one redundant form output sum vector and
one redundant form output carry vector. If the results of the
multiplications performed by multipliers 505 and 506 are
independent of each other, they can bypass blocks 508 and 510. For
example, the 27 less-significant bits of the vectors from
multiplier 505 can bypass blocks 508 and 510 via the AC and AS
links shown in FIG. 21. (In FIG. 21 the upper circle AC is
connected to the lower circle AC; the upper circle AS is connected
to the lower circle AS; the upper circle BC is connected to the
lower circle BC; and the upper circle BS is connected to the lower
circle BS. These connections are shown in this way to reduce
crowding in the drawing and also to reduce the amount of possibly
confusing line cross-over in the drawing.) Similarly, the 27
less-significant bits of the vectors from multiplier 506 can bypass
blocks 508 and 510 via the BC and BS links shown in FIG. 21. The
outputs of the AC link are applied to one selectable set of the
inputs to multiplexer circuitry ("mux") 660c. The 27
less-significant bits of the carry vector output by block 510 are
applied to the other selectable set of inputs to mux 660c. The
outputs of the AS link are applied to one selectable set of the
inputs to mux 660s. The 27 less-significant bits of the sum vector
output by block 510 are applied to the other selectable set of
inputs to mux 660s. The outputs of the BC link are applied to one
selectable set of inputs to mux 662c. The 27 more-significant bits
of the carry vector output by block 510 are applied to the other
selectable set of inputs to max 662c. The outputs of the BS link
are applied to one selectable set of inputs to mux 662s. The 27
more-significant bits of the sum vector output by block 510 are
applied to the other selectable set of inputs to mux 662s. CRAM 664
controls which selectable set of inputs each of muxes 660 and 662
selects as its outputs (i.e., muxes 660 and 662 output either their
AC, AS, BC, and BS inputs, or they output their inputs from block
510). Adder 511 can be controlled to operate either as two separate
(nominally) 27-bit adders (separated by the dotted line shown in
FIG. 21), or as one (nominally) 54-bit adder (not separated at the
dotted line). Such separation of adder 511 into two parts just
means that arithmetic carry from right to left across the dotted
line is disabled. If no such separation is desired, arithmetic
carry across the dotted line is enabled. CRAM 666 controls whether
carry across the dotted line is enabled or disabled, and therefore
whether adder 511 operates as one large adder or as two, separate,
smaller adders. This portion of the specification thus describes
how multipliers 505 and 506 may be operated to produce two
multiplication results that can be output separately (although in
this embodiment each of those separate products is limited to 27
bits). The AC, AS, BC, and BS links are used to bypass blocks 508
and 510 in this mode of operation.
[0121] If rather than being kept separate, the results of the
multiplications performed by multipliers 505 and 506 need to be
added together, those results preferably go through compressor 508
before going to the final adder stage farther downstream. The
result from multiplier 506 can be (optionally) left-shifted by 18
bit positions going through elements 507 before going into block
508. This controllably selectable left-shift option provides the
flexibility of different multiplication configurations. The
above-described 27.times.27 multiplication is one example in which
compressor 508 is used to add together the results from multipliers
505 and 506, with the results from multiplier 506 first being
shifted left 18 bit positions by operation of elements 507. Another
example of use of compressor 508 (and elements 507) is 36.times.18
multiplication as will now be described beginning in the next
paragraph.
[0122] Lines 1 and 2 of FIG. 22 show a 36.times.18 multiplication
to be performed. This is multiplication two 18-bit data words x2
and x1 by 18-bit data word y1. The partial multiplication of x1 by
y1 (line 3 in FIG. 22) is performed by multiplier 505 operating in
the above-described 18.times.18 multiplication mode. The other
partial multiplication of x2 by y1 (line 4 of FIG. 22) is performed
in multiplier 506 in the same DSP block 500 as includes the
just-mentioned multiplier 505. The left-shift by 18 bit positions
shown in line 4 of FIG. 22 is performed by controllable shift
elements 507 operating on the outputs of multiplier 506. Compressor
508 combines the outputs of multiplier 505 and elements 507 to
produce redundant form sum and carry vectors indicative of the
results shown in line 5 of FIG. 22. This result is, of course, the
54-product of the 36.times.18 multiplication of x2 and x1 by
y1.
[0123] If no further operations need to be performed on the outputs
of compressor 508, further 4:2 compressor circuitry 510 can be
bypassed/disabled, and only final, carry-propagate adder ("CPA")
circuitry 511 is enabled in order to send the final result to
output register circuitry 512.
[0124] An advantage of the FIGS. 11-22 design is to efficiently
combine multiple DSP blocks or modules 500 to perform more complex
operations. The purpose of 4:2 compressor circuitry 510 is to take
the two outputs from compressor 508, along with one or two vectors
from another DSP module, and compress all of that data into two
redundant form vectors, which become the inputs to final CPA 512.
Two scenarios in which DSP modules 500 can be combined together by
taking advantage of the inter-module connections sum_in/sum_out and
car_in/car_out (where car is an abbreviation for carry) will now be
described.
[0125] In the first scenario, multiplexer circuitry 516 (controlled
by CRAM circuitry 670) is used to select the result "out" through
the output chaining path. (This path extends from output registers
512, through mux 680 (controlled by CRAM 682), mux 516, and mux
517c (controlled by CRAM 672c).) In this way, two or more DSP
modules 500 can be chained together through the car_in/car_out
connections between adjacent ones of those modules 500. For
example, the module 500 on the left as viewed in FIG. 23 can be
used to add to its own internally generated results the results
from the other adjacent module 500 on the right in that FIG.
[0126] A second scenario involving use of direct, dedicated,
inter-DSP-block connections like sum_in/sum_out and car_in/car_out
between adjacent DSP blocks 500 is as follows. In this scenario,
sum and carry vectors from element 508 in one DSP block (instead of
the final result from the output chaining including the output
registers 512 in that block) are connected through the
sum_in/sum_out and car_in/car_out tracks between adjacent blocks
500. This again allows a DSP block 500 (e.g., like the one on the
left in FIG. 23) to combine (e.g., add together) its own internally
generated results and results from another DSP block (e.g., the DSP
block on the right in FIG. 23). However, this second scenario can
result in better FMAX performance (maximum speed of operation)
because the data do not have to go through the delay of blocks 510,
511, and 512 in one DSP block 500 (e.g., the one on the right in
FIG. 23) before combining with the data in the next DSP block 500
(e.g., the one on the left in FIG. 23). However, this type of
connection may be limited to chaining only a relatively small
number (e.g., two) DSP blocks 500 together. Any applications that
utilize only such a small number of DSP blocks 500 (e.g., two such
blocks) may be able to achieve a better result with this approach.
Controllable shift elements 509s and 509c (like other similar shift
elements throughout this disclosure, and controlled by CRAM
elements 674s and 674c, respectively) provide data shifting
capability to steer (select) the correct data from the current DSP
block 500 (e.g., on the right in FIG. 23) to the next DSP block 500
(e.g., on the left in FIG. 23).
[0127] Again, it will be remembered that muxes 517s and 517c
(controlled by CRAM bits 672s and 672c, respectively (or possibly
in tandem by one such CRAM bit)) are the so-called redundancy
muxes. As just a brief reminder, these muxes allow data to flow via
the sum_in/sum_out and car_in/car_out tracks from either (a) the
DSP block 500 immediately adjacent a receiving DSP block 500, or
(b) the DSP block 500 that is one block away from the receiving DSP
block (e.g., in the event that the immediately adjacent DSP block
is defective and therefore must be functionally replaced by the
slightly more distant DSP block). In the representative
configuration shown in FIG. 24, this redundancy capability means
that "receiving" DSP block 500c can receive its sum_in and car_in
signals from either the sum_out and car_out signals of immediately
adjacent DSP block 500b, or from the sum_out and car_out signals of
slightly more distant DSP block 500c.
[0128] To expressly state what should already be apparent, FIGS. 23
and 24 (like FIGS. 1, 6, 8, 10, 30 and 32) illustrate the point
that an integrated circuit in accordance with this disclosure
includes multiple instances (i.e., multiple identical or
substantially identical repetitions) of DSP circuit blocks or
modules (e.g., 100, 500, or the like). Thus terms like "DSP circuit
block," "DSP block," "DSP module," etc., as used herein refer to an
assembly of circuit elements that is suitable for identical or
substantially identical repetition on an integrated circuit. At
least some of the circuitry in such a DSP circuit block is
preferably dedicated (i.e., hard-wired or at least partly
hard-wired) to performing particular functions such as
multiplication, redundant form sum and carry vector compression,
addition, etc. Adjacent (or at least nearby) ones of these DSP
circuit blocks may have "dedicated" connections between them.
Examples of such dedicated inter-DSP-block connections are the
sum_out/sum_in and car_out/car_in connections between adjacent
blocks. Such dedicated inter-DSP connections are preferably usable
only for conveying signals between DSP blocks.
[0129] Returning once again to FIG. 11 or FIG. 24, mux circuitry
514 (controlled by CRAM element 676) can be used to select the
result from the previous operating (clock) cycle and combine it
with (e.g., add it to) the result of the current operating (clock)
cycle. (Each operating cycle typically takes place during one
respective cycle of the clock signal that is used to clock the
registers employed in the circuitry.) This accumulation path is
optional in every DSP module 500.
[0130] Elements 504 and 513 are systolic delay registers. They can
be used, for example, for 18-bit, systolic, finite-impulse-response
("FIR"), digital filtering applications. The principles of systolic
FIR filters are described in the next several paragraphs.
[0131] FIR filters are widely used in DSP applications. The basic
structure of a FIR filter is a plurality of parallel
multiplications with addition of the resulting products. A FIR
filter operation can be represented by the equation shown in FIG.
25. In the FIG. 25 equation x[n-i-1] represents the sequence
(succession) of input digital data samples (each sample being,
e.g., an 18-bit binary word presented in parallel), c[i] represents
the filter coefficients (each coefficient being, e.g., an 18-bit
binary word presented in parallel), and k-1 is the number of taps
that the filter has.
[0132] Circuitry 800 for implementing (in so-called direct form)
the FIR filter equation shown in FIG. 25 is shown in FIG. 26. In
this FIG., elements 810 are multiplier circuitries, elements 820
are delay circuitries (each of which delays the data sample applied
to it by the time duration of any one sample in the input sample
stream x[n]), and elements 830 are adder circuitries. (The
just-mentioned "time duration" is also typically the "period" of
the clock signal that is used to clock registers used throughout
the circuitry.)
[0133] FIG. 26 shows that a FIR filter 800 in the direct form
includes a large adder tree, which is formed by chaining adders
830. Depending on the number of taps (equal to the number of
multipliers minus 1) and the input sample size (the number of bits
in each input sample x[n]), the delay through adders 830 can become
quite large. In order to overcome this possible performance issue,
the systolic form 800' of the filter shown in FIG. 27 can be used.
In particular, FIG. 27 shows that additional delay circuit elements
812 and 822 are placed per tap to increase the performance of the
overall circuitry, albeit at the cost of increased latency (i.e.,
overall delay through the circuit from input x[n] to output y[n]).
More particularly, FIG. 27 shows an additional delay element 822 in
the sample stream between each tap; and a matching, additional
delay element 812 in the product-summing stream, also between each
tap. In other words, additional delay element 812-1 "matches" or
"corresponds to" additional delay element 822-1, additional delay
element 812-2 matches or corresponds to additional delay element
822-2, and so on. All of delay elements 812, 820, and 822
preferably have the same delay (e.g., the period of one cycle in
the clock signal that is used to clock registers throughout the
circuitry, which "period" is also preferably equal to the duration
of each successive sample in the input same stream x[n]).
[0134] Additional delay elements 812 and 822 may sometimes be
referred to as systolic registers. Delay elements 812 break the
otherwise long chain of adders 830 down into individual adders.
Each adder 830 is therefore followed by a register 812 which
registers the sum produced by that adder before passing that sum on
to the next adder. This avoids having to operate the circuitry at
the relatively slow speed that would otherwise be required for data
to propagate all the way through a long adder chain in one
operating (clock) cycle of the circuitry. Systolic filter 800'
allows a faster clock to be used because each adder 830 can
complete its addition operation relatively quickly, and the result
of that addition is not needed or used by the next adder until the
next clock cycle. Delay elements 822 are added to keep the
propagation of data samples through the data sample path
synchronized with the delayed propagation of product sums through
the product-sum-accumulation path.
[0135] Although systolic filter 800' tends to have greater latency
(overall delay through the filter from input x[n] to availability
of the corresponding output y[n]) than a direct form filter 800
with the same number of taps (due to the additional delays 812 and
822 in filter 800'), other circuitry (e.g., on an integrated
circuit with the FIR filter) can benefit greatly (in terms of
operating speed) from having a faster clock signal, which filter
800' permits the circuitry to have.
[0136] To help better relate the immediately preceding discussion
of systolic FIR filters like 800' in FIG. 27 to the earlier
discussion of DSP blocks like 500, FIG. 27 is redrawn (with some
small changes) in FIG. 28. The only differences between FIG. 28 and
FIG. 27 are: (1) FIG. 28 is reoriented to better correlate with
other FIGS. like FIG. 11; (2) FIG. 28 is made less general by
omitting the FIG. 27 depiction of possible additional taps between
multipliers 810-2 and 810-(k-1), and by showing only four
immediately adjacent taps; and (3) FIG. 28 is made more general (or
more modular) by showing another adder 830-1 in the lower
right-hand corner.
[0137] Assume that each of multipliers 810 in FIG. 28 is an
18.times.18 multiplier (e.g., like a multiplier 505 or 506 in FIG.
11). It can then be efficient to perform the operations associated
with each pair of two adjacent taps in FIG. 28 in one DSP block 500
constructed as shown, for example, in FIG. 11. The following
discussion shows how this can be done.
[0138] FIG. 29 shows circuitry 800'' with the same capabilities as
circuitry 800' in FIG. 28. In other words, FIG. 29 is an
alternative implementation of the FIG. 28 circuitry that is
functionally equivalent to the FIG. 28 circuitry. In FIG. 29, the
function of systolic registers 812-1 (FIG. 28) is performed by
systolic registers 504p-1, 504q-1, and 513-1 (although in the FIG.
29 example, systolic registers 513-1 are not used). Basically, the
delay produced by systolic registers 812-1 between adders 830-1 and
830-2 in FIG. 28 is moved (in FIG. 29) upstream along all paths
that feed the addition performed by adder 830-1. This enables the
additions performed by adders 830-1 and 830-2 in FIG. 28 to be
performed in one combined adder 508/510/511-1 in FIG. 29 (i.e.,
with no systolic delay like 812-1 being needed within composite
adder 508/510/511-1). Similarly, the delay produced by systolic
registers 812-3 in FIG. 28 is moved upstream to systolic registers
504p-3, 504q-3, and 513-3 in FIG. 29. Again, this enables the
additions performed by adders 830-3 and 830-4 in FIG. 28 to be
performed in one combined adder 508/510/511-3 in FIG. 29 (i.e.,
again with no systolic delay like 812-3 needed within composite
adder 508/510/511-3).
[0139] With the foregoing as a starting point, the chain-dotted
lines 500e and 500f in FIG. 29 can now be explained. Chain-dotted
line 500e surrounds the elements in FIG. 29 that can be implemented
by the circuitry of one DSP block 500 having the construction
shown, for example, in FIG. 11. In particular, the following
equalities of reference numbers show how the elements inside line
500e are implemented by the elements in a representative DSP block
500 as in FIGS. 11: 810-2=506; 504p-1=504p; 504q-1=504q; 810-1=505;
508/510/511-1=508, 510, and 511; 512-1=512; 513-3=513.
[0140] Chain-dotted line 500f surrounds the elements in FIG. 29
that can be implemented by the circuitry of another (one) DSP block
500, again having the construction shown, for example, in FIG. 11
and able to directly receive via its car_in inputs the outputs of
the systolic registers 513 in DSP block 500e. Again, the following
equalities of reference numbers show how the elements inside line
500f are implemented by the elements in this second ("another")
instance of FIG. 11 DSP block circuitry 500: 810-4=506;
504p-3=504p; 504q-3=504q; 810-3=505; 508/510/511-3=508, 510, and
511; 512-3=512; 513-5=513.
[0141] The delay performed by elements 820 and 822 in FIG. 29 may
be performed by register circuitry that is outside of DSP blocks
like 500e and 500f, and that can appropriately delay the
application of successive input data words x[n] to the various
inputs to DSP blocks 500e and 500f. (Alternatively, some or all
elements 820 and 822 may be implemented by registers that are
included in DSP blocks that are otherwise like blocks 500e and
500f.)
[0142] The preceding paragraphs demonstrate how DSP blocks 500
constructed as shown, for example, in FIG. 11 can be used to
efficiently implement systolic FIR filters if desired. Although
FIG. 29 shows implementing a systolic FIR filter having four taps
in two adjacent DSP blocks 500e and 500f, it will be understood
that systolic FIR filters having any desired number of taps can be
implemented in the same general way in the appropriate number of
adjacent DSP blocks 500. For example, a 6-tap systolic FIR filter
can be implemented in three adjacent DSP blocks 500; an 8-tap
systolic FIR filter can be implemented in four adjacent DSP blocks
500; and so on. In each case, two adjacent taps of the systolic FIR
filter ("a pair of adjacent taps" or "a pair of two adjacent taps")
are implemented in one DSP block 500, and each DSP block 500
implements a respective one pair of such mutually exclusive pairs
of taps. The DSP blocks are ordered in the same way that the taps
in the pairs are ordered, all of which pairing and ordering of taps
(from 1 to 4), and similar ordering of DSP blocks 500 (from e to
f), are shown illustratively in FIG. 29.
[0143] We return now to consideration of the broader range of
capabilities of DSP blocks 500. The following is a list of the
major features that a single DSP block or module 500 can
support:
1. Three 9.times.9 multiplications. 2. One 18.times.18
multiplication. 3. Two partial 18.times.18 multiplications. 4. One
27.times.27 multiplication. 5. One 27.times.27 multiplication with
accumulation. 6. One 26.times.22 multiplication with one 25-bit
pre-adder. 7. One 36.times.18 multiplication. 8. One sum of two
18.times.18 multiplications. 9. One sum of two 18.times.18
multiplications with accumulation. 10. One sum of two 18.times.18
multiplications with pre-adders. 11. One sum of square difference.
12. (a*b)+c, (a, b=18 bits each, c=36 bits). 13. One sum of two
18-bit systolic FIR filter taps. 14. One sum of two 18-bit systolic
FIR filter taps with pre-adders.
[0144] In the case of three 9.times.9 multiplications (item 1
above), two 18.times.18 multipliers (blocks 505 and 506) can be
configured as four 9.times.9 multipliers. However, the number of
outputs (56) is only sufficient to support three 9.times.9
multiplication results (18 bits+18 bits+18 bits=54) bits). The same
restriction applies to the case of two partial 18.times.18
multiplications (item 3) above, where only 28 bits (out of 36 bits)
of the result of each 18 bit multiplication can be routed to the
output (out) of the DSP module.
[0145] In the application of one 26.times.22 multiplication with
one 25-bit pre-adder (item 6 above), the restriction comes from the
number of inputs to DSP block 500. Even though DSP block 500 can
support up to 27-bit multiplication, the number of dynamic inputs
is only 72, which is only enough to support (25+25)*22.
[0146] One sum of square difference (item 11 above) can be
implemented by enabling the pre-adders 502 and applying the results
of the pre-adders to both inputs of the multipliers (505 and 506).
Thus in this case the result of (ax-ay).sup.2+(bx-by).sup.2 can be
obtained at the output.
[0147] The following is a list of features that can be supported
when multiple DSP blocks or modules 500 are combined together:
a. Three 18.times.18 multiplications in two DSP modules. b. Sum of
four 18.times.18 multiplications. c. Sum of two 27.times.27
multiplications. d. Sum of two 36.times.18 multiplications. e.
Complex 18.times.18 multiplication (i.e., (a+bi)*(c+di), where a,
b, c, and d are each 18 bits). f. 36.times.36 multiplication. g.
72.times.18 multiplication. h. Complex 18.times.25 multiplication
in three DSP modules (i.e., (a+bi)*(c+di), where a and b are each
18 bits, and c and d are each 25 bits). i. Complex 27.times.27
multiplication (i.e., similar to item e above, except that a, b, c,
and d are each 27 bits). j. 54.times.54 multiplication. k. Sum of
two 18.times.18 multiplications with output chaining. l. One
27.times.27 multiplication with output chaining.
[0148] One DSP module 500 is not sufficient to support the outputs
of two 18.times.18 multiplications. However, two DSP modules 500
are sufficient to support the outputs of three 18.times.18
multiplications with one result distributed between the two
modules.
[0149] For applications such as sum of two 36.times.18
multiplications (item d above), and 72.times.18 multiplication
(item g above), two DSP blocks 500 are connected through the
dedicated, direct, inter-module sum and carry tracks. The following
discussion contains more information about how these connections
work.
[0150] FIG. 30 shows two adjacent DSP blocks 500h and 500i. Each of
these DSP blocks can be a respective instance of circuitry like
that shown, for example, in FIG. 11. Some of the connections shown
in FIG. 30 are emphasized by being drawn using heavier (thicker)
lines. These include the direct, dedicated, inter-DSP-block
connections that are used when the two depicted DSP blocks are
being used together to perform certain functions that exceed the
capacity of one DSP block 500 alone. An example is a 36.times.36
multiplication (item f) above. FIG. 31 shows how such a
multiplication can be performed using the FIG. 30 circuitry. FIGS.
30 and 31 will now be further described for the example of the
36.times.36 multiplication shown in FIG. 31.
[0151] Line 1 of FIG. 31 shows the 36-bit multiplicand for the
36.times.36 multiplication to be performed. Line 2 of FIG. 31 shows
the 36-bit multiplier for that multiplication. The less significant
18-bit sub-word x1 of the multiplicand and the less significant
18-bit sub-word y1 of the multiplier are applied to multiplier 505
in DSP block 500h for multiplication by that multiplier to produce
the partial product shown in line 3 of FIG. 31. The more
significant 18-bit sub-word x2 of the multiplicand and y1 are
applied to multiplier 506 in DSP block 500h for multiplication by
that multiplier to produce the partial product shown in line 4 of
FIG. 31. Elements 507s and 507c in DSP block 500h are used to shift
the redundant form sum and carry vectors output by multiplier 506
in DSP block 500h 18 bits to the left (to increase their arithmetic
significance by 18 bit positions relative to the redundant form sum
and carry vectors output by multiplier 505 in block 500h). This
18-bit left-shift of partial product x2*y1 is shown graphically in
line 4 of FIG. 31. Compressor 508 in DSP block 500h compresses the
four redundant form vectors it receives down to two redundant form
vectors, i.e., the 54-bit sum vector shown in line 5 of FIG. 31,
and the 54-bit carry vector shown in line 6 of FIG. 31.
[0152] The 18-bit portions of the data in lines 5 and 6 in FIG. 31
to the right of the dotted vertical line in that FIG. remain in DSP
block 500h. In particular, these 18 least significant bits pass
down through compressor 510 in block 500h to CPA 511 in that block.
CPA 511 in block 500h converts this redundant form data to the
least significant 18 bits of the non-redundant-form (or normal
form) final result shown in line 9 of FIG. 31. DSP block 500h can
output these 18 final result bits via its OUT leads.
[0153] The more significant 36-bit portions of the data in lines 5
and 6 in FIG. 31 (to the left of the dotted vertical line) are
routed from the outputs of compressor 508 in DSP block 500h to DSP
block 500i. In particular, the 36 more significant sum bits output
by compressor 508 in block 500h are passed through controllable
shifter element 509s in block 500h to shift that data 18 bit
positions to the right to decrease the data's arithmetic
significance by that amount as it enters DSP block 500i via the
direct, dedicated, sum_out/sum_in connections (conductors or
tracks) from block 500h to block 500i. (Mux circuitry 517s in block
500h is part of this routing.) Similarly, the 36 more significant
carry bits output by compressor 508 in block 500h are passed
through controllable shifter 509c in block 500h to shift that data
18 bit positions to the right to decrease its significance as it
enters DSP block 500i via the car_out/car_in connections
(conductors or tracks) from block 500h to block 500i. (Mux
circuitry 517c in block 500h is part of this routing.) The
redundant form data that DSP block 500i thus receives from DSP
block 500h is routed to compressor circuitry 510 in block 500i.
(Mux circuitry 514 in block 500i is part of this routing.)
[0154] The more significant 18-bit sub-word y2 of the multiplier
and x1 are applied to multiplier 505 in DSP block 500i for
multiplication by that multiplier to produce the partial product
shown in line 7 of FIG. 31. Similarly, y2 and x2 are applied to
multiplier 506 in DSP block 500i for multiplication by that
multiplier to produce the partial product shown in line 8 of FIG.
31. Controllable shifter circuit elements 507s and 507c in DSP
block 500i are used to shift the redundant form sum and carry
vector signals output by multiplier circuitry 506 in DSP block 500i
18 bit positions to the left (to increase their arithmetic
significance by 18 bit positions relative to the redundant form sum
and carry vector signals output by multiplier 505 in block 500i).
This 18-bit left-shift of partial product x2*y2 relative to partial
product y2*x1 is shown graphically in line 8 of FIG. 31.
[0155] Compressor 508 in DSP block 500i reduces the four redundant
form partial product vectors it receives from elements 505, 506,
and 507 in block 500i to two redundant form vectors indicative of
the sum of partial products y2*x1 and x2*y2. These two redundant
form vectors from compressor 508 in DSP block 500i are applied to
compressor 510 in block 500i. Compressor 510 in block 500i
compresses the four redundant form vectors that it receives (two
from compressor 508 in block 500i, and two from block 500h via the
sum_out/sum_in and car_out/car_in tracks from block 500h to block
500i as described earlier) down to two redundant form vectors
indicative of the sum of (1) partial product y2*x1, (2) partial
product x2*y2, and (3) the 36 more significant bits of the sum of
(a) partial product x1*y1 and (b) partial product x2*y1. The two
redundant form vectors output by compressor 510 in DSP block 500i
are applied to CPA 511 in block 500i, which combines those two
vectors to the 54 more significant bits of the non-redundant-form
(or normal form) final result shown to the left of the dotted line
in line 9 of FIG. 31. These 54 more significant final result bits
can be output by DSP block 500i via its OUT leads.
[0156] The foregoing demonstrates how two DSP blocks 500h and 500i
can be used to perform a full 36.times.36 multiplication, with the
18 less significant bits of the final product being output by DSP
block 500h, and with the 54 more significant bits of the final
product being output by DSP block 500i.
[0157] For applications such as sum of four 18.times.18
multiplications (items b above) and sum of two 27.times.27
multiplications (item c above), the result can be achieved by
combining two DSP modules 500 through either the sum and carry
direct, dedicated, inter-DSP-module connections (i.e.,
sum_out/sum_in and car_out/car_in) as illustratively shown and
described above in connection with FIGS. 30 and 31, or through
output chaining connections as illustratively shown in FIG. 32 and
described below.
[0158] FIG. 32 shows several DSP blocks 500-1 through 500-7 that
are preferably adjacent to one another (e.g., in the relative
positions to one another that are shown in FIG. 32) on an
integrated circuit. Each of these DSP blocks can be a respective
instance of circuitry like that shown, for example, in FIG. 11.
Some of the connections shown in FIG. 32 are emphasized by being
drawn using heavier (thicker) lines. These include connections
between adjacent DSP blocks 500 that can be used when it is desired
to chain the final output (OUT) of one DSP block 500 directly into
intermediate circuitry of another adjacent DSP block for
combination with signal processing that is in progress in the
second ("another" or "receiving") DSP block. Like the
earlier-described sum_out/sum_in and car_out/car_in connections,
these output chaining connections are preferably direct, dedicated,
inter-DSP-block connections. This means, for example, that these
direct, dedicated connections do not make use of the more
general-purpose signal routing circuitry on the integrated circuit
that includes the DSP blocks. In other words, these connections are
dedicated to only conveying signals between instances of DSP
circuit blocks (e.g., DSP blocks 500). In the illustrative
embodiments being described, the direct, dedicated output chaining
connections are made (in part) via the previously described
car_out/car_in connections from each DSP block 500 to the next
adjacent DSP block 500.
[0159] In the illustrative embodiment shown in FIG. 32, the
emphasized output chaining routing from DSP block 500-1 to DSP
block 500-2 extends from the node in block 500-1 that can receive
signals from the output registers 512 in block 500-1. (This may be
referred to as the OUT node of block 500-1.) This routing extends
through (optionally usable) systolic registers 513 in block 500-1,
mux circuitries 516 and 517c in block 500-1, and onto the car_out
leads of block 500-1 that connect to the car_in leads of block
500-2. As previously described, signals on the car_in leads of a
DSP block 500 are applied (inter alia) to the compressor circuitry
510 in that DSP block. Thus the output chaining routing from DSP
block 500-1 to DSP block 500-2 includes application of the car_in
signals received by block 500-2 from block 500-1 to the compressor
510 in block 500-2. Compressor 510 in block 500-2 is thus the
"intermediate circuitry" in block 500-2 that can receive outputs of
block 500-1 chained into block 500-2 for combination with the
results of other signal processing performed in block 500-2 by
elements upstream from compressor 510 in block 500-2. Note that
this output chaining does not convey outputs from source block
500-1 to the most upstream inputs (e.g., ax, ay, bx, by) of
receiving block 500-2. On the contrary, the output chaining from
source block 500-1 enters receiving block 500-2 at an intermediate
point in the circuitry of receiving block 500-2. This leaves all
the most-upstream inputs (ax, ay, bx, by) of receiving block 500-2
available for other (non-output-chained) inputs. It also leaves
other upstream circuitry (e.g., elements 501, 502, 503, 504, 505,
506, 507, and 508) of receiving block 500-2 available for
processing those other (non-output-chained) inputs. Then the
results of that processing of other inputs by the upstream elements
in receiving block 500-2 can be combined with the output-chained
inputs (received from the OUT node of source block 500-1 via the
car_out/car_in link from block 500-1 to block 500-2) in the more
downstream elements in block 500-2 (i.e., starting with the
intermediate circuit element compressor 510 in DSP block
500-2).
[0160] FIG. 32 shows (by additional emphasized (thicker) connection
lines) that the output chaining can continue from DSP block 500-2
to the next adjacent DSP block 500-3 (not actually shown in FIG.
32). In block 500-2 this continued output chaining (nominally)
starts at the OUT node of block 500-2 and passes through that
block's elements 513, 516, and 517c to the car_out output leads of
block 500-2 (which are directly connected to the car_in inputs to
block 500-3).
[0161] FIG. 32 still further shows (using the illustrative example
of DSP block 500-7) that output chaining may include accumulation
of successive results (respectively produced in successive clock
cycles of the circuitry). Thus in block 500-7 the OUT node signals
are routed back (via mux circuitry 514) to compressor 510 in block
500-7. This allows the elements 510, 511, and 512 in block 500-7 to
combine (e.g., add together) and accumulate from each clock cycle
to the next clock cycle (1) outputs of the more upstream circuitry
(e.g., 505 and 506) in block 500-7, (2) outputs chained in from
another, adjacent, "source" DSP block 500-6 (not actually shown),
and (3) the final outputs produced by DSP block 500-7 during the
immediately preceding clock cycle. FIG. 32 also shows that the
results of such accumulator operation by block 500-7 can be "output
chained" on to the next adjacent DSP block 500-8 (not shown) via
elements 513, 516, and 517c in block 500-7.
[0162] It is emphasized that although FIG. 32 shows the output
chaining going through the systolic registers 513 in each DSP block
500, those registers can either be used to delay the chained
signals or not, as desired by the user of the circuitry. In other
words, actual use of registers 513 for added signal delay is
optional (see, again, FIG. 16).
[0163] As an example of use of the circuit arrangements like those
shown and emphasized in FIG. 32, a sum of four 18.times.18
multiplications (item b above) can be obtained at the output of DSP
block 500-2, with each of DSP blocks 500-1 and 500-2 set to
calculate a sum of two 18.times.18 multiplications.
[0164] In general (with continued reference to FIG. 32), any number
of "adjacent" DSP modules 500 can be connected together via the
output chaining path. ("Adjacent" is quoted in the preceding
sentence because the redundancy muxes 517 in each DSP module 500
allow a module to be skipped or bypassed (e.g., if it is not
functioning properly). This principle applies to all references to
"adjacent" DSP blocks or modules throughout this disclosure.) The
summation of all the results can be obtained at the last DSP module
500 at the end of the chain. Note again that the path between the
OUT node and mux circuitry 514 in DSP block 500-7 (for example) can
be enabled to incorporate the accumulation feature with the output
chaining.
[0165] Still referring to FIG. 32, complex 18.times.18
multiplication (item e above) or complex 27.times.27 multiplication
(item i above) can be implemented using two sets of sum of two
18.times.18 or 27.times.27 multiplications (item b or item c
above), with one set doing the addition (i.e., of ad and be in the
imaginary part adi+bci of the complex product), and with the other
set doing the subtraction (i.e., of ac-bd in the real part of the
complex product). Complex 18.times.25 multiplication (item h above)
can be implemented by using four DSP modules 500 with the same
setting as complex 27.times.27 multiplication. 54.times.54
multiplication (item j above) can be implemented with four DSP
modules 500 utilizing both the sum and carry direct connections and
the output chaining direct connections.
[0166] Although FIGS. like FIG. 11 and FIG. 21 show redundancy
multiplexers like 517s and 517c at or near outputs of DSP blocks
500, such redundancy muxes can instead be at or near inputs to such
DSP blocks. For example, FIG. 33 shows portions of two adjacent DSP
blocks 500a' and 500b' illustrating this possible modification. In
particular, instead of redundancy muxes 517s and 517c being at or
near outputs of DSP block 500a', those muxes have been moved to the
input side of DSP block 500b'. In all other respects, DSP blocks
like 500a' and 500b' can be like the DSP blocks 500 shown and
described elsewhere in this specification.
[0167] As another illustration of how the locations and/or order of
various elements can change without departing from the scope and
spirit of this disclosure, systolic register circuitry 513 can be
moved to any other suitable point along the output chaining path in
(e.g.) FIG. 21. For example, systolic register circuitry 513 and
the associated bypass elements 680 and 682 can be moved to a point
along the output chaining path that is just prior to that path
entering the next downstream DSP block.
[0168] Recapitulating some aspects of the foregoing, and also in
some respects extending what has been said, a digital signal
processing ("DSP") circuit block (e.g., 500 in FIG. 11; 500b in
FIG. 24) may include a first multiplier circuit (e.g., 506) that
can perform a first N-bit by N-bit ("N.times.N") multiplication
(e.g., an 18.times.18 multiplication); a second multiplier circuit
(e.g., 505) that can perform a second N.times.N multiplication;
first shifter circuitry (e.g., 507c/507s) for shifting outputs of
the first multiplier circuit by a selectable one of (1) zero bit
positions and (2) N bit positions toward greater arithmetic
significance; first compressor circuitry (e.g., 508) for additively
combining outputs of the first shifter circuitry and the second
multiplier circuit; circuitry (e.g., SUM_OUT/CAR_OUT leads) for
routing outputs of the first compressor circuitry to a first other
DSP circuit block (e.g., 500c in FIG. 24) that is adjacent to the
DSP circuit block; and second compressor circuitry (e.g., 510) for
additively combining outputs of the first compressor circuitry and
outputs received from a second other DSP circuit block (e.g., 500a
in FIG. 24) that is adjacent to the DSP circuit block.
[0169] A DSP circuit block as recapitulated above may further
include adder circuitry (e.g., 511) for additively combining
outputs of the second compressor circuitry. Such adder circuitry
may comprise carry-propagate adder circuitry.
[0170] A DSP circuit block as recapitulated above may further
include first register circuitry (e.g., 512) for registering
outputs of the adder circuitry.
[0171] A DSP circuit block as recapitulated above may further
include first, second, third, and fourth pluralities of data signal
input leads (e.g., ax, ay, bx, by) for supplying inputs for
processing by the multiplier circuits; and pre-adder circuitry
(e.g., 502) for additively combining the inputs from at least two
of the pluralities of data signal input leads.
[0172] A DSP circuit block as recapitulated above may further
include circuitry (e.g., 503) for selecting outputs of the
pre-adder circuitry as an alternative to inputs from at least one
of the pluralities of data signal input leads for processing by at
least one of the multiplier circuits.
[0173] A DSP circuit block as recapitulated above may further
include selectively usable systolic register circuitry (e.g., 504p,
504q) in input data signal paths to one of the multiplier
circuits.
[0174] A DSP circuit block as recapitulated above may further
include selectively usable systolic register circuitry (e.g., 513)
in output data signal paths for data output by the second
compressor circuitry.
[0175] A DSP circuit block as recapitulated above may further
include circuitry (e.g., 512) for registering signals indicative of
outputs of the second compressor circuitry; and circuitry (e.g.,
514) for selectively applying outputs of the circuitry for
registering to inputs of the second compressor circuitry for
additive combination by the second compressor circuitry with the
outputs of the first compressor circuitry.
[0176] In a DSP circuit block as recapitulated above the circuitry
for routing may include second shifter circuitry (e.g., 509s/509c)
for shifting the outputs of the first compressor circuitry by a
selectable one of (1) zero bit positions and (2) N bit positions
toward lesser arithmetic significance.
[0177] In a DSP circuit block as recapitulated above the circuitry
for routing may include alternative routing circuitry (e.g.,
SUM_IN/CAR_IN leads) for routing to the first other DSP circuit
block signals output by circuitry for routing in the second other
DSP circuit block.
[0178] In a DSP circuit block as recapitulated above the circuitry
for routing may include alternative routing circuitry (e.g., output
chaining leads) for routing to the first other DSP circuit block
signals indicative of outputs of the second compressor
circuitry.
[0179] In a DSP circuit block as recapitulated immediately above
the alternative routing circuitry may include systolic register
circuitry (e.g., 513) that is selectively usable for the signals
indicative of outputs of the second compressor circuitry.
[0180] In a DSP circuit block as recapitulated above the second
multiplier circuit may alternatively perform (1) a 0.5N-bit by
M-bit ("0.5N.times.M") multiplication (e.g., a 9.times.27
multiplication), and (2) a 0.5N bit by N bit ("0.5N.times.N")
multiplication (e.g., a 9.times.18 multiplication), where M is
equal to 1.5N (e.g., M is 27 when N is 18).
[0181] A DSP circuit block as recapitulated immediately above may
further include circuitry (e.g., 503) for selectively routing bits
of an M-bit multiplicand data word and an M-bit multiplier data
word to the first and second multiplier circuits so that the first
multiplier circuit can multiply the N most significant bits of the
multiplicand data word by the N most significant bits of the
multiplier data word to produce a first partial product data word,
and so that the second multiplier circuit can multiply (1) the M
bits of the multiplicand data word by the 0.5N least significant
bits of the multiplier data word to produce a second partial
product data word, and (2) the N most significant bits of the
multiplier data word by the 0.5N least significant bits of the
multiplicand data word to produce a third partial product data
word.
[0182] In a DSP circuit block as recapitulated immediately above
the second multiplier circuit may include third compressor
circuitry (e.g., 652) for additively combining the second and third
partial product data words, with bits of the third partial product
data word being increased in arithmetic significance by 0.5N bit
positions relative to bits of the second partial product data
word.
[0183] In a DSP circuit block as recapitulated above the circuitry
for routing may be dedicated to only conveying signals between
instances of the DSP circuit block.
[0184] In a DSP circuit block as recapitulated above the outputs
received from a second other DSP block may be received via
circuitry (e.g., SUM_IN/CAR_IN leads) that is dedicated to only
conveying signals between instances of the DSP circuit block.
[0185] Recapitulating some other aspects of the foregoing, and also
in some respects extending what has been said, circuitry (e.g.,
FIG. 17) that is usable to perform a selectable one of (1) one
M-bit by M-bit ("M.times.M") multiplication (e.g., a 27.times.27
multiplication) and (2) two N-bit by N-bit ("N.times.N")
multiplications (e.g., two 18.times.18 multiplications), where M is
equal to 1.5N, may include N.times.N multiplier circuitry (e.g.
506); 0.5N.times.M multiplier circuitry (e.g., 650a); 0.5N.times.N
multiplier circuitry (e.g., 650b); first circuitry (e.g., 652) for
additively combining outputs of the 0.5N.times.M multiplier
circuitry and outputs of the 0.5N.times.N multiplier circuitry;
shifter circuitry (e.g., 507c/507s) for shifting outputs of the
N.times.N multiplier circuitry by a selectable one of (1) zero bits
positions and (2) N bit positions toward greater arithmetic
significance; and second circuitry (e.g., 508) for additively
combining outputs of the shifter circuitry and outputs of the first
circuitry.
[0186] In circuitry as recapitulated above N may be equal to 18 and
M may be equal to 27.
[0187] Circuitry as recapitulated above may further include first,
second, third, and fourth pluralities of data signal input leads
(e.g. any four of d, c, ax, ay, bx, by in FIG. 11); and routing
circuitry (e.g., 503) for controllably routing signals from each of
the pluralities to selectable ones of inputs to the multiplier
circuitries. Each of the pluralities may consist of N data signal
input leads.
[0188] In circuitry as recapitulated immediately above the routing
circuitry may include first, second, third, and fourth N-bit
multiplexers (e.g., 630r, 630s, 630u, 630w), each of which is
controllable to route N bits from any one of the pluralities to a
respective one of (1) multiplicand inputs to the N.times.N
multiplier circuitry, (2) multiplier inputs to the N.times.N
multiplier circuitry; (3) N less significant multiplicand inputs to
the 0.5N.times.M multiplier circuitry, and (4) multiplicand inputs
to the 0.5N.times.N multiplier circuitry; and first, second, and
third 0.5N-bit multiplexers (e.g., 630t, 630v, 630x), each of which
is controllable to route 0.5N bits from any of at least some of the
pluralities to a respective one of (1) 0.5N more significant
multiplicand inputs to the 0.5N.times.M multiplier circuitry, (2)
multiplier inputs to the 0.5N.times.M multiplier circuitry, and (3)
multiplier inputs to the 0.5N.times.N multiplier circuitry.
[0189] Recapitulating some still other aspects of the foregoing,
and also in some respects extending what has been said, a digital
signal processing ("DSP") circuit block (e.g., 500 in FIG. 11; 500b
in FIG. 24) may include first means (e.g., 506) for performing a
first N-bit by N-bit ("N.times.N") multiplication (e.g., an
18.times.18 multiplication); second means (e.g., 505 in FIG. 11;
FIG. 17) for performing a selectable one of (1) a second N.times.N
multiplication and (2) a 0.5N-bit by M-bit ("0.5N.times.M")
multiplication and a 0.5N-bit by N-bit ("0.5N.times.N")
multiplication, where M is equal to 1.5N (e.g., M is 27 when N is
18); third means (e.g., 507c/507s) for shifting outputs of the
first means by a selectable one of (1) zero bit positions and (2) N
bit positions toward greater arithmetic significance; fourth means
(e.g., 508) for additively combining outputs of the second and
third means; fifth means (e.g., SUM_OUT/CAR_OUT leads) for routing
outputs of the fourth means to a first other DSP circuit block
(e.g., 500c in FIG. 24); and sixth means (e.g., 510) for additively
combining outputs of the fourth means and outputs received from a
second other DSP circuit block (e.g., 500a in FIG. 24).
[0190] In such a DSP circuit block the second means may include
seventh means (e.g., 650a) for performing the 0.5N.times.M
multiplication; eighth means (e.g., 650b) for performing the
0.5N.times.N multiplication; and ninth means (e.g., 652) for
additively combining outputs of the seventh and eighth means, with
bits of the outputs of the eighth means being increased in
arithmetic significance by 0.5N bit positions relative to bits of
the outputs of the seventh means.
[0191] A DSP circuit block such as is here being recapitulated may
further include tenth means (e.g., 630s) for selectively routing
the N most significant bits of an M-bit multiplicand data word to
multiplicand inputs to the first means; eleventh means (e.g., 630r)
for selectively routing the N most significant bits of an M-bit
multiplier data word to multiplier inputs to the first means;
twelfth means (e.g., 630v and 630w) for selectively routing the
M-bit multiplicand to the multiplicand inputs to the seventh means;
thirteenth means (e.g., 630x) for selectively routing the 0.5N
least significant bits of the multiplier data word to multiplier
inputs to the seventh means; fourteenth means (e.g., 630u) for
selectively routing the N most significant bits of the multiplier
data word to multiplicand inputs to the eighth means; and fifteenth
means (e.g., 630t) for selectively routing the 0.5N least
significant bits of the multiplicand data word to the multiplier
inputs to the eighth means.
[0192] A DSP circuit block such as is here being recapitulated may
further include sixteenth means (e.g., ax, ay, bx, and by leads)
for receiving multiple pluralities of data input signals for use as
inputs to the first and second means; seventeenth means (e.g., 502)
for selectively pre-adding at least some of the data input signals;
and eighteenth means (e.g., 503) for selectively applying outputs
of the seventeenth means to selected inputs to at least one of the
first and second means.
[0193] A DSP circuit block such as is here being recapitulated may
further include nineteenth means (e.g., 504p, 504q) for selectively
giving inputs to the second means a systolic register delay.
[0194] In a DSP circuit block such as is here being recapitulated
the fifth means may include twentieth means (e.g., 509s/509c) for
shifting the outputs of the fourth means by a selectable one of (1)
zero bit positions and (2) N bit positions toward lesser arithmetic
significance.
[0195] In a DSP circuit block such as is here being recapitulated
the fifth means may include twenty-first means (e.g., 516) for
selectively alternatively routing output signals indicative of
outputs of the sixth means to the first other DSP circuit
block.
[0196] In a DSP circuit block such as is here being recapitulated
the twenty-first means may include twenty-second means (e.g., 513)
for selectively giving the output signals indicative of outputs of
the sixth means a systolic register delay.
[0197] In a DSP circuit block such as is here being recapitulated
the fifth means may include twenty-third means (e.g., SUM_IN/CAR_IN
leads) for alternatively routing signals from the second other DSP
circuit block to the first other DSP circuit block.
[0198] In a DSP circuit block such as is here being recapitulated
the sixth means may include twenty-fourth means (e.g., 512) for
registering output signals indicative of outputs of the sixth
means; and twenty-fifth means (e.g., 514) for applying outputs of
the twenty-fourth means to inputs to the sixth means.
[0199] It will be understood that the foregoing is only
illustrative of the principles of the invention, and that various
modifications can be made by those skilled in the art without
departing from the scope and spirit of the invention. For example,
the use of 18.times.18 multipliers 200 is only illustrative, and
other sizes of multipliers can be used instead, if desired. It will
also be understood that the FIGS. herein like FIGS. 1-4, 6, 8, and
10 show electronic circuitry in which most or all of the circuit
elements (like 200, 202, 210, 220, 260, etc.) are separate elements
(typically integrated together on an integrated circuit), each of
which elements is dedicated (e.g., hard-wired to at least a
significant extent) to performing the function(s) described for it
herein by operating on electrical input signals to produce other
electrical output signals. These electrical signals may sometimes
be referred to as data, bits, vectors, "1", "0", values,
multiplicand, multiplier, product, partial product, sum, or the
like; but in all cases they are in fact actual electrical signals
representing the specified information. Similarly, certain of the
electrical circuit elements shown herein may sometimes be referred
to by the operations they perform (e.g., "multiplier 200a"), but in
all cases these elements are actual electrical circuitry (e.g.,
multiplier circuitry) performing those operations. The same is true
for references to DSP circuitry, DSP operations, or the like. All
such operations are performed by electrical circuitry operating on
electrical signals as mentioned above.
[0200] In certain of the appended claims, words like first, second,
third, fourth, etc., may be used. This is done only to provide
unique identifiers for various claim elements and not for any other
purpose. For example, these words are not intended to imply
anything about how elements are ordered or used. Also, a reference
to a higher numbered element (e.g., the "fourth" element) does not
by itself imply that all lower numbered elements (e.g., the first
through third elements) are also present in a particular claim.
* * * * *